A survey of recommender systems for energy efficiency in buildings: Principles, challenges and prospects
Yassine Himeur, Abdullah Alsalemi, Ayman Al-Kababji, Faycal Bensaali, Abbes Amira, Christos Sardianos, George Dimitrakopoulos, Iraklis Varlamis
AA SURVEY OF RECOMMENDER SYSTEMS FOR ENERGYEFFICIENCY IN BUILDINGS : P
RINCIPLES , CHALLENGES ANDPROSPECTS
A P
REPRINT
Yassine Himeur ∗ , Abdullah Alsalemi, Ayman Al-Kababji, Faycal Bensaali Department of Electrical EngineeringQatar UniversityDoha, Qatar [email protected];[email protected];[email protected];[email protected]
Abbes Amira
Institute of Artificial IntelligenceDe Montfort UniversityLeicester, United Kingdom [email protected]
Christos Sardianos, George Dimitrakopoulos, Iraklis Varlamis
Department of Informatics and TelematicsHarokopio University of AthensAthens, Greece [email protected];[email protected];[email protected]
February 16, 2021 A BSTRACT
Recommender systems have significantly developed in recent years in parallel with the witnessedadvancements in both internet of things (IoT) and artificial intelligence (AI) technologies. Accord-ingly, as a consequence of IoT and AI, multiple forms of data are incorporated in these systems,e.g. social, implicit, local and personal information, which can help in improving recommender sys-tems’ performance and widen their applicability to traverse different disciplines. On the other side,energy efficiency in the building sector is becoming a hot research topic, in which recommendersystems play a major role by promoting energy saving behavior and reducing carbon emissions.However, the deployment of the recommendation frameworks in buildings still needs more investi-gations to identify the current challenges and issues, where their solutions are the keys to enable thepervasiveness of research findings, and therefore, ensure a large-scale adoption of this technology.Accordingly, this paper presents, to the best of the authors’ knowledge, the first timely and com-prehensive reference for energy-efficiency recommendation systems through (i) surveying existingrecommender systems for energy saving in buildings; (ii) discussing their evolution; (iii) providingan original taxonomy of these systems based on specified criteria, including the nature of the rec-ommender engine, its objective, computing platforms, evaluation metrics and incentive measures;and (iv) conducting an in-depth, critical analysis to identify their limitations and unsolved issues.The derived challenges and areas of future implementation could effectively guide the energy re-search community to improve the energy-efficiency in buildings and reduce the cost of developedrecommender systems-based solutions. K eywords Recommender systems · energy efficiency · evaluation metrics · artificial intelligence · explainablerecommender systems · visualization. ∗ Information Fusion, 2021 a r X i v : . [ c s . I R ] F e b PREPRINT - F
EBRUARY
16, 2021
The use of energy has been increasing exponentially through the last few years around the world. Specifically, thebuilding sector alone consumes more than 40% of the global energy produced worldwide [1, 2]. This consumptionis expected to increase by 1.3% per year on average from 2018 to 2050 in organization for economic cooperationand development (OECD) countries (e.g. USA, Europe, Australia, etc.), while this rate will be more than 2% fornon-OECD countries (e.g. Middle East, China, Russia, etc.) [3]. Therefore, experts naturally assume that the rise ofpopulation and quality of life in various regions will result in a growing need for electricity-consuming devices andindividualized equipment, and hence, an increasing energy consumption rate [4, 5].In order to alleviate this issue, recent research and development projects and initiatives have been focused on devel-oping nearly zero energy buildings (nZEB) in the last decade, which incorporate renewable and sustainable energyresources and energy management systems [6]. However, these kind of measures could not be supported in all coun-tries around the globe due to its high deployment cost [7, 8]. Consequently, finding other cost-effective or no-costenergy saving solutions became the core of interest for the building energy community, especially those based on theuse of information and communication technologies (ICT) [9]. One of these challenging approaches is behavioralchange, which allows end-users to polish their energy consumption behaviors and trim their wasted energy withoutinvesting more time and elbow grease, but only by using recommender algorithms, artificial intelligence (AI) toolsand already used smartphones [10, 11, 12]. To this regard, energy providers, policy makers and end-users in thebuilding sector have become progressively aware of the importance of behavioral change in promoting energy savingand reducing carbon emissions [13, 14]. In this context, an increasing number of literature, projects and commercialproducts have recently arisen to explore the research interest of sustainable behavior change, explicitly to address therelation between attitudes in order to improve energy consumption behavior [15]. This is also due to the widespreaduse of AI, Internet of things (IoT) devices and other ICT tools, which have a positive impact on raising end-users’awareness, shaping their attitudes towards energy saving and boosting their achievements [16, 17].While most of the research efforts have been conducted towards developing and improving new technologies andmaterials that reduce wasted energy and promote energy saving, human-related aspects, especially those related toend-user’s behavior [18] have received less attention. Therefore, strategies and objectives must be set in order toshape the behaviors of buildings’ end-users and owners [19]. This can be achieved through developing context-aware recommender systems [20] that combine the knowledge of AI, behavioral analytics and human decision-makingprocesses, to implement powerful behavioral change support systems [21, 22], in which recommendations could beeasily embedded into daily behaviors to reach an effective energy saving level. In this regard, changing daily behaviorof end-users has become a key challenge [23, 24]. This challenge requires training and awareness exercises, incentiverecommendations and feedback assessments for inducing a permanent change.Despite the success of recommender systems in different research and development applications (e.g. in healthcare,online shopping, movies, music, travel plans, etc.), there is still room for more research to improve their performance,especially in the energy saving domain. According to our knowledge, no work yet has been dedicated to the surveyof challenges, difficulties and future perspectives of energy saving recommender systems. As a result, there are stillopen questions, in the energy research community, about the reliability and efficacy of recommendation systems. Toalleviate theses issues, we provide this survey article that performs an in-depth, critical analysis of energy savingrecommender systems for buildings. More specifically, this paper identifies the open issues and critical challenges thatimpede the development of effective energy recommender systems that incorporate human-in-the-loop, by promotingenergy efficiency behaviors and reducing carbon emissions.To that end, a taxonomy of energy recommender systems is presented. The currently available frameworks are de-scribed and the factors that impact the efficacy of current implementation are discussed. Such factors refer to thenature of developed recommender algorithms, the computing platforms used to implement them, their objectives, theincentive measures utilized to motivate end-users and the evaluation metrics that fit the context of energy. Movingforward, a critical analysis and discussion is conducted to identify the limitations and difficulties encountered whendeveloping energy recommender systems. Finally, we derive and decipher the issues that remain unresolved and attractan increasing research interest along with the hottest research directions for recommendation systems’ performanceimprovement. To summarize, the main contribution axes of the paper could be outlined as follows:• Propose the first review framework of recommender systems for energy efficiency in buildings.• Conduct a novel taxonomy of existing energy efficiency recommender systems through analyzing the na-ture of the different components used to build a recommender framework, e.g. (i) objective of the recom-mendations, (ii) methodology and algorithm of choice utilized by the recommender engine, (iii) computingplatforms, and (iv) evaluation metrics and incentive measures.2
PREPRINT - F
EBRUARY
16, 2021• Perform in-depth, critical analysis of the existing frameworks to identify the current challenges and difficultiesthat remain unresolved.• Provide insights about the future orientations that could be targeted to overcome existing energy efficiencyrecommender systems’ issues, to improve their quality and facilitate their applicability.The remainder of this paper is organized as follows. Section 2 briefly explains the methodology that we followedin order to identify the works related to the development of energy saving recommender systems. Section 3 beginswith an introduction to recommender systems, and then focuses on the analysis of human-driven energy efficiencyframeworks of buildings, using recommendation systems for triggering and maintaining this behavioral change. Ad-ditionally, it summarizes the objectives of such systems, methodologies and algorithms they use, computing platformsand evaluation metrics. Moving forward, Section 4 conducts in-depth and critical analyses of existing energy recom-mender systems by discussing their limitations and issues. Following, Section 5 presents current challenges and futureorientations that should attract the attention of R&D communities in the near and far future. Lastly, Section 6 derivesthe final conclusions.
In order to perform our literature review, we based our methodology on the techniques presented in [25]. Identifyingthe need for a review is of equal importance to the results of the review. The need for this review derives from the factthat lots of systems and solutions have been developed in the field of energy efficiency for buildings, which add to thevariety of the research field. Our study reveals that there is not yet a systematic review that explains all the steps: fromthe conception of an energy saving solution, to the delivery to end-users. By conducting this review, the followingquestions will be answered:1. Why have recommender systems gained significant attention for energy efficiency in buildings?2. What are the main research directions that existing energy saving recommender system frameworks followed?3. What are the main objectives and which methodologies were used to achieve them?We performed our bibliometric research under the perspective of a narrative review. Studies related to the use ofrecommendations for improving energy efficiency and promoting energy savings in buildings have been searched.Our search took place through the Scopus database from 2000 to 2020. The following terms have been searchedin titles, abstracts and keywords: “recommendation”, “recommenders”, “recommender systems”, “energy saving”,“energy efficiency”, “buildings”, “behaviour”.A search in Scopus , in the title, abstract and keyword fields returned
283 articles , which are broadly organized in three major research directions , as we explain in the following paragraphs and depict in Fig. 1:1. Recommendations for enhancing buildings energy efficiency2. Intelligent systems that promote energy saving in buildings3. Recommender systems that put humans in the center of the decision making process for energy efficiency, ineach of the previous cases or in bothA large group of works focuses on energy efficient buildings by design. Such “high-performance” buildings employenergy optimization techniques by including natural ventilation, thermal storage and optimal window size and place-ment. The number of works in this field is vast, but most of them are slightly related to recommender systems andhuman behavior. So we suggest readers to consult a few survey works in this domain [26, 27].Another major group of works focuses on the use of intelligent systems for monitoring and reducing the buildings’unnecessary energy waste. The survey of Boodi et al. [28] summarizes the state-of-the-art works in building energymanagement system (BEMS) and distinguishes three types of models that combine environmental conditions, energyprices, comfort criteria and occupancy prediction in order to optimize the heating, ventilation and air conditioning(HVAC) systems operation: white box, black box and gray box models. More work on intelligent systems for energyefficiency in buildings can be found in a few more surveys [29, 30, 31].The third group of works employ recommendation systems and algorithms, usually as a complement to the previoustwo approaches (i.e. on energy efficiency and smart buildings). The final decision is always on the human, who plays TITLE-ABS-KEY ( ( “energy saving” OR “energy efficiency” OR energy ) AND ( “recommendation” OR “recommender” OR“recommender systems” OR “recommendation systems” ) AND behaviour AND buildings ) PREPRINT - F
EBRUARY
16, 2021Figure 1: The impact of energy efficiency in buildings in bringing human-in-the-loop.a vital role in the efficacy of the proposed solution. The recommendations in this group are either targeted to buildingowners (mainly for large public or commercial buildings) or to the occupants of residential buildings. The maindifference between the two, is that in the former case, recommendations refer to an energy plan or an energy savingstrategy that can be used to balance between saving and comfort, whereas in the latter one, the recommendations areabout energy saving actions (e.g. device turn-off, or work shifting) that can have an immediate impact on the building’senergy consumption.As depicted in Fig. 1, the focus of our survey is on recommender systems for energy efficiency and more specificallyon the intersection of the three aforementioned domains. In the following, we examine various aspects of systemsthat combine intelligent systems and action recommendations to improve energy efficiency and convert conventionalbuildings into smart ones. The section that follows begins with an overview of recommender systems and answers thelast question of our survey methodology by presenting the objectives and methodologies used to achieve them.
According to the 2012 ACM Computing Classification System , recommender systems are categorized as informationsystems that focus on information retrieval tasks. Under this prism, it is now very common for more and morescenario-specific applications to adopt various kinds of recommender systems to serve their needs and goals. Althoughthey have originally been applied online for content personalization based on users’ explicit or implicit preferences[32, 33], they soon have been extended to a wide range of different real-world applications [34], from place andpeople recommendations based on location [35] to action recommendations for reshaping energy profiles [36, 37].In addition, due to the high demand of personalization in most of real-life scenarios, various approaches of adoptingrecommender systems have been proposed based on the type of decision making the system has to support and thegoal these recommendations have to meet. Although all these approaches still share the same logic as before (i.e.recommend items to users), they go beyond content personalization and consequently increase the requirements fromrecommended item, which apart from matching a specific user’s interests, also have to be novel, profitable, feasible interms of the user context, generated at the right place and moment [38].Figure 2: Phases of recommendation process. PREPRINT - F
EBRUARY
16, 2021The used approach for implementing a recommender system significantly depends on the goal and applicability of thesystem. As depicted in Fig. 2, the main phases that each recommendation process consists of are:1.
Information collection phase : In this phase, all the necessary explicit or implicit user information is col-lected in order to profile the targeted users. Depending on the type of the used recommender system, theinformation needed for user profiling might differ in volume or attributes, but in minimum it has to cover thecase scenario for which the recommender system is used. It is critical, in this phase, to recognize and preparethis information, which will be used for system training and fine-tuning in the next phase. The efficiencyand impact of the resulting recommendations rely on the utilized algorithm, but also depend on the quality oftraining data.2.
Learning phase : In this phase, the system extracts the most representative features and trains the modelthat best identifies and quantifies the relationship among the users and the “items” that the recommendationengine will create recommendations for.3.
Prediction and recommendation phase : In the third and last phase of the recommendation process, thesystem predicts the “unknown” values of the user-to-items preferences using the pre-trained model, andranks the items that are most likely to fit users’ preferences. As a result, it adds the top ranked items in thelist of recommendations that are to be presented to the user. Of course, filters can be employed to rule outitems that do not match the user context, and more criteria can be used to improve the ranking of items to berecommended [39].In following sections, we present a taxonomy of energy saving recommender systems, in which we describe state-of-the-art research frameworks with the aim of identifying the most prominent and recent advances of this technology.Fig. 3 illustrates a taxonomy of energy saving recommender systems that is introduced based on different parameters,including the objective of the recommender system, the recommendation methodology, the computing platforms, theevaluation metrics, and the employed measures to encourage end-users to adopt energy saving behavior.
The analysis performed in this article focuses on the third group of research works that we identified through theliterature survey. As we explained earlier, the group comprises two main subgroups, each one targeting a differentaudience, and thus, having different objectives.
O1. Strategy recommenders:
The objective of such systems is to recommend the best strategy for each case, either itis a building, an energy consumption forecast model, or an operational energy setup for HVAC or lights. For example,[40] proposes a building energy efficiency recommender system that relies on case-based reasoning (CBR). In a similardirection, [41] discusses a recommender system for selecting the operational light intensity that satisfies user’s comfortthat is suitable for both user activity and energy saving. Finally, [42] presents a generalized structure for forecastingbuilding energy trends based on building specifications. The solution is mostly targeted to energy prosumers andallows them to predict the total energy consumption of a building by choosing the right building profile.
O2. Action recommenders:
These systems are tailored to the everyday needs of building occupants and their mainobjective is to recommend actions that minimise the energy footprint of occupants . The actions can either assumestatic users and inelastic needs, or build on their flexibility to move around the building or postpone their needs at alater time (e.g. using the laundry machine after hours). For example, [43] develops a Resource-oriented rule-basedengine that generates advice for energy savings in the form of real-time alerts or logged incidents for monitoringpurposes. On the other side, [44] focuses on commercial buildings and distinguishes two types of recommendations,which both aim to maximize the space usage: recommendations for occupants to move from one space to another,and recommendations for occupants to shift their schedule related to the building. In a more recent work [45], authorspresent a recommender system for reducing energy consumption in commercial buildings. They employ human-in-the-loop methodologies and utilize deep reinforcement learning in order to learn actions with energy saving capability andactively deliver recommendations to building end-users. Once again, the recommendations have to balance betweenuser comfort and energy efficiency, but the final decision lies in the end-users’ hands. For example, ReViCEE [46]analyzes historic energy usage fingerprints and provides individual and collaborative recommendations that balancebetween comfort and efficient usage of energy.
In this section, we overview recommendation systems methodologies widely used in the building energy saving field.5
PREPRINT - F
EBRUARY
16, 2021Figure 3: Taxonomy of energy saving recommender systems.
M1. Case-based recommender systems are in essence rule-based systems, which recommend actions for one or moreend-users through handling every end-user separately. Explicitly, specific energy usage habits and preferences areexamined with reference to an ensemble of rules/heuristics and predetermined decisions that initiate, if achieved, theassociated energy efficiency acts.In [47], a rule-based recommender system is implemented to effectively learn the energy consumption patterns andinterests of end-users, and hence, allows them to independently promote energy efficiency. In this regard, a frequent-sequential data mining model is deployed to extract the characteristics of consumers. Similarly, in [48], a pairwiseassociation rule-based algorithm is proposed for drawing the collective preferences of groups of end-users. The re-sulting recommender system is able to provide personalised suggestions to users without requiring a complex ratingapproach. While in [40], the authors propose a case-based reasoning recommendation system, which is built based onsystem knowledge (i.e. cases) representing the historical energy consumption actions in order to shape the end-userbehavior towards an energy efficiency comportment. Explicitly, the system is able to suggest energy efficiency actionsto end-users at every different moment of the day. This is done by analyzing their energy usage footprints and compar-ing them with the ones already stored in the knowledge base. In this line, in order to identify the similar behaviors atevery time-stamp, a k-Nearest Neighbor (KNN) approach is deployed, while a support vector machine (SVM)-basedmethod is utilized to optimize the weighting parameters of every example. Moving forward, an expert system is thenemployed, which includes a set of ad-hoc rules for ensuring the application of the developed scheme to the case underconsideration. Recently, in [49], pattern mining techniques are used for creating appliance usage profiles that consideruser time context. Then, they filter out the most prominent time-appliance patterns that indicate the ideal behavior and6
PREPRINT - F
EBRUARY
16, 2021when they detect an appliance usage that deviates from the typical behavior, they send a recommendation notificationto the end-user to interact accordingly with the device.
M2. Collaborative filtering methods assume that the ensemble of end-users select from a closed set of actions (oritems). These explicit or implicit choices are used to filter (predict) the preference of any end-user for any action in theset of available actions (or items). Therefore, the recommended actions to each end-user represent the most preferredto him/her (or those having the highest estimated rating) [50] or those generated from the same group he/she pertainsto [51]. In this line, energy efficiency recommender systems, using collaborative filtering, utilize various intelligentagents, which interact proficiently and dynamically identify the end-users’ preferences. This helps in promotingenergy saving actions to consumers by providing them with personalized recommendations that appropriately fit theirinterests.For example, in [52], appliances’ consumption data of a specific household are analyzed before predicting the ratinglevels of different energy usage plans and identifying the related user preferences for every plan. Moving forward, afiltering (prediction) model is used to allow users to choose suitable consumption plans with adequate tariffs. In thesame way, the authors in [53] opt to generate energy efficiency recommendations based on a dual-step procedure, i.e.extracting features and triggering tailored recommendations. Explicitly, at the first stage, a matrix representation isadopted to adapt user preferences with reference to appliances usage. Following, a collaborative filtering approachis employed to detect similar users and generate personalized recommended actions using KNN clustering. While in[46], the authors design the ReViCEE recommender system, which delivers tailored recommendations to end-usersat a university campus building in Singapore, helping them curtail their electricity usage. Accordingly, ReViCEElearns end-users’ interests by analyzing historic energy usage fingerprints. In this regard, individual and collaborativepreferences related to the use of lights are captured from current consumption patterns before triggering a set ofrecommendations to grant end-users the best compromise between visual comfort and energy saving.
M3. Context-aware recommender systems aim to produce more pertinent recommendations which are adjusted tothe particular contextual circumstances of the end-user [54]. This can be based on the detection of historic energyconsumption patterns and their underlying context, which can then be used to develop rule-based recommendationsaiming to achieve end-user’s satisfaction [55, 56]. They usually require a longer interaction with end-users with thesystem in order to collect a larger amount of data that helps in a better adaptation of the generated recommendationsto the specific contextual situation of the consumer [57].The authors in [58] introduce a context-aware recommendation system based on (i) the analysis of power consumptionfootprints in various contexts; (ii) the maintenance of a base of historic formulated recommendations to avoid repli-cated recommendations; and (iii) a social survey for evaluating the recommendations efficiency, in which 47 usersadopted the recommender system suggestions and provided feedback on their efficiency. The empirical assessmentillustrates that when the recommendations are chosen randomly and in large numbers, they can overflow users andmay have a detrimental impact on the end-users.
M4. Rasch-based models represent a psychometric paradigm that is generally utilized to analyze the user’s responsesto a specific set of recommendations [59]. It aims at identifying the best compromise between the user’s behavior andthe ability to implement the generated recommendations. In this line, a Rasch-based recommender system relies onperforming a Rasch analysis, which explains the probability of the end-user to perform a particular recommendationas a function of the end-user’s ability and the recommendation’s difficulty [60].For example, the authors in [61] investigate to what extent Rasch-based recommendations could help in reducing theend-user’ efforts, improving the system assistance while increasing the choice satisfaction and leading to the promotionof energy efficiency behaviors in buildings. To this end, a Rasch-based recommendation system is developed, whereup to 79 energy-efficiency recommendations are generated to assist end-users in making correct energy usage actions,enhancing system support, collecting their feedback and rating their satisfaction.
M5. Probabilistic relational models (PRM) can be used to capture the knowledge hidden in the energy consump-tion data in a probabilistic manner, which represents the probability of an action to match certain usage patterns orpreferences [62]. PRMs replace the user-item preference matrix with a relational database with “Users” and “Items”being the main entities and real transactions being the captured relationship between the two. When a transactionbetween a user and an item is recorded, the probabilities for users and items with similar attribute values increase. Inthis essence, probabilistic relational paradigms are developed for predicting end-user’s preferences and habits, wheretailored recommendations are derived to motivate end-users to reduce their wasted energy [63].7
PREPRINT - F
EBRUARY
16, 2021In [64], the authors record and analyze historical energy usage footprints in a work space via the use of a continuousMarkov chain model. The model focuses on investigating time-series energy data using a multi-objective program-ming scheme. Next, a tailored energy usage advice is drawn and each end-user is notified to establish energy savingmeasures and support the adoption of renewable energy solutions. Similarly, in [45], a recommender system aimingat optimizing power usage in commercial buildings is introduced using human-in-the-loop model. Moving forward,the energy saving approach for a building is implemented and modeled based on a Markov decision paradigm, anda deep reinforcement learning scheme is implemented to learn energy saving recommendations, and engage con-sumers in effective sustainable behaviors. The implemented system helps in learning end-users’ energy consumptionactions/habits with a good accuracy, and hence, results in providing end-users with useful and personalized recommen-dations. Finally, further experiments are conducted to employ a feedback rating procedure to evaluate user satisfactionand identify the best recommendations.
M6. Fusion-based models rely on the analysis of different kinds of data, such as energy consumption footprints,ambient conditions (i.e. temperature, humidity, light, and luminosity), outdoor weather information and user prefer-ences/habits, for producing better and well-timed recommendations [65]. The so-called fusion-based recommendationsystems adopt data fusion approaches, which either collect and analyze different kinds of data representations from dis-tinct sources before making decisions [66] or include various sub-recommenders and aggregate their recommendationoutputs, thus, building a recommendation ensemble [67, 68].For example, in [69], authors introduce a multi-modal embedding fusion-based recommendation system that combinesinformation from multiple sources and modalities. In [70], a hybrid recommender system is proposed, which is basedon the assumption that the end-user choice is generally impacted by its direct (and even indirect) friends’ preferences.The system fuses different kinds of data (e.g. social data, score, and review patterns) and trains a preference predictionmodel, using a joint-representation learning process, to extract the best recommendations. In [71], authors incorporateadditional information from the consumer’s social trust network as well as actionable semantic-domain knowledge,in order to improve the recommendations accuracy and increase their coverage. In the same manner, in [72], by ex-ploiting different social data sources (produced by the Internet, e.g. consumer profiles, social relationships, behaviors,preferences, etc.), a recommendation system using social data fusion is proposed. Explicitly, it aims to utilize socialdata fusion for identifying similar consumers, and hence, updating each consumer rating of recommended actionsusing similar consumers.In [73], a multi-level fusion-based recommender system is developed to produce collaborator recommendations. Itfuses deep learning and biased random walk models to provide tailored recommendations for possible end-users hav-ing the same preferences. Following, in [74], a recommender system is introduced to support sustainable greenhousemanagement in buildings using multi-sensor data aggregation based system. Specifically, contextual information,mathematical formulations and experts’ knowledge are used and fused to help in generating more effective recom-mendations.
M7. Deep learning-based recommender systems, have gained significant attention recently in various research topics,including visual recognition, healthcare, fraud detection, natural language processing, etc. [75]. Their use is extendeddue to their remarkable performance in many learning tasks, and additionally because of their interesting ability tolearn characteristic representations from the ground up. The impact of deep learning is widespread as well to otherresearch issues, in which it demonstrates its efficiency to retrieve information and trigger recommendations. Evidently,the field of deep learning in recommender system is flourishing [76, 77].In [78], with the aim of addressing the gap in collaborative filtering-based methods, a deep learning model is adopted.In effect, collaborative filtering systems are seriously suffering from the cold start issue, especially with the absence ofhistoric data about the users and their energy consumption preferences. Moreover, the latent parameters learnt by thesesystems are naturally linear. To that end, deep learning is employed, where embeddings are deployed to represent usersand their preferences, and thus, to allow the learning of non-linear latent parameters. This approach better alleviates thecold start issue, since information about users and their preferences is embedded in the deep learning model. In [79],a collaborative filtering approach, called DeepMF, which combines deep neural networks with matrix factorization isintroduced for improving both the predictions of end-users’ preferences and the provided recommendations. In thiscontext, DeepMF applies an iterative refinement of a matrix factorization paradigm based on multilayer architecture,in which the acquired knowledge from a layer is used sequentially as input for the following layer, and so on.
M8. Classical optimization is proposed in addition to learning-based recommender systems to achieve an optimalenergy usage in households and other kinds of buildings. Indeed, an essential part of the literature review is basedon classical optimization techniques, which can (i) provide relevant recommendations to both end-users and energyproviders; and (ii) reduce wasted energy automatically through controlling energy demand and electrical devices8
PREPRINT - F
EBRUARY
16, 2021[80, 81]. For example, in [82], the authors propose an energy optimization approach, which aims to predict theamount of energy used by the heating and cooling systems in a set of commercial or institutional buildings. Following,the potential impacts of various energy saving measures based on parameter optimization are investigated beforerecommending tailored actions to optimize energy consumption. Similarly, in [83], Rocha et al. discuss the potentialof energy saving in buildings using efficient energy and policy measures. Accordingly, an optimization model thatmimics a smart building energy management system is developed through the aggregation of the decisions on heatingand cooling systems operations with decisions on energy demand. Moving forward, an optimization and ontology-driven multi-agent recommender system is introduced in [84] to reduce wasted energy. It can monitor and optimizeenergy usage within an integrated home/building and/or microgrid systems using different renewable energy resourcesand controllable loads. In this regard, several agents are developed and integrated together with the aim of improvingtheir cooperation and optimizing the operation strategy of the whole energy system.In [85], in order to optimize energy saving and guarantee a perfect thermal comfort of the end-users, the uncertaintiesdue to outdoor weather conditions, building parameters and human behaviors are thoroughly modeled. Following,an adaptive economic dispatch approach is introduced, which is based on conducting a thermal comfort managementprocess using a two-step algorithm. Similarly, in [86], aiming at reducing energy demand uncertainties and energybills in households and small public buildings, a two-step energy monitoring framework is proposed. Specifically,uncertainties due to energy demand variations and prediction errors of renewable generation are detected before gen-erating appropriate recommendations to reduce wasted energy. Moreover, in [87, 88], energy usage optimizationis conducted by forecasting the heat and moisture transfer, which directly affect the indoor climate and the overallthermal performance of buildings.In [89], reducing wasted energy in various buildings is ensured by optimizing the consumed energy taking into consid-eration the abrupt changes produced by the rooftop solar generation and the real-time price of energy. Accordingly, anovel parameter called the load criticality rate has been deployed, which represents the threshold value applied by eachbuilding occupants to their power consumption. Furthermore, the energy reduction task is considered as a stochastic,multi-objective optimization issue. While in [90], an analytical model that describes thermal dynamic characteristicsof district heating networks in buildings is developed to optimize energy consumption while keeping an acceptablecomfort level of the end-users.
Aiming at bridging the gap between development and implementation of energy efficiency recommender systems, thissection presents the main computing solutions that can be utilized, as a standalone solution or in an integration ofvarious ones, to implement these systems.
P1. Cloud computing relies on the use of cloud infrastructure principles to provide reliable and scalable methodsto solve resource-intensive computational issues [91]. The employment of cloud-based services and solutions inhome automation and building energy monitoring has been widely and globally popular. The cloud computing modelfacilitates a broad variety of processing applications, exploits the potentials of IoT and leverages IoT processingrestrictions by moving the most demanding parts such as deep learning algorithms [92] to the cloud. The mainchallenge for cloud computing solutions remains to be the privacy of IoT data [93] that are transferred and processedon the cloud platform.
P2. Edge computing receives increasing attention although transmitting data to the cloud for processing has becomea central topic in recent decades, pushing cloud computing as a prevailing model in computing [94]. In effect, therapid growth in the amount of devices and data traffic in the age of IoT puts major hurdles on the capacity-limitedInternet and on unregulated service delays. Through utilizing cloud storage alone, it becomes impossible to fulfill thedelay-critical and context-aware service specifications of IoT apps. Met with these problems, computing paradigmsare moving from clustered cloud computing to dispersed edge computing. Edge computing allows some of the dataprocessing to be done on the device (i.e. the edge) to lift some of the burden off the cloud server [95, 96, 97] andguarantee privacy, grace to the decentralised processing.
P3. Fog computing has been recently used for developing distributed, low-energy recommender systems based on IoTarchitectures. Their main applications are in the domains of healthcare [98], banking [99], or information brokerage[100]. Despite the many advantages of fog computing, which comprise low latency, privacy, uninterrupted service andlocation awareness, there is still no application that combines fog computing with energy efficient recommendations.9
PREPRINT - F
EBRUARY
16, 2021
P4. Hybrid edge-cloud approaches stand between cloud and edge computing, and their primary aim is to increasethe flexibility of IoT systems by moving data processing from the edge to the cloud and vice-versa, depending on thesystem constraints [101]. For example, when a small dataset is at use, hybrid systems decide to offload it to the edgedevice for rapid processing and for reducing the communication overhead, while in more demanding situations, dataare pushed to the cloud in order to get more efficient results. Also, employing an edge-cloud architecture can helpdistribute computations between the edge and cloud for performance optimization [102]. This flexibility of hybridapproaches creates new, out-of-the-box possibilities for more powerful yet resource-efficient IoT solutions.
Modelling a recommender system that fits the need of the business/initiative, encompasses an evaluation phase thattests the recommender’s capabilities to the limits. A recommender system suggests items (or actions) to the users,based on their own expected preferences. Several metrics have been devised for this purpose that allow system per-formance evaluation in predicting and providing sensible recommendations that fit a given scenario. Among the longlist of metrics that can be used for the evaluation of provided recommendations [103, 104], we highlight those metricsthat we see relevant to an energy efficiency recommendation system.
E1. Rating accuracy measures the deviation between the predicted and the actual ratings assigned by a user to eachrecommended item. The simplest and most popular error are
Mean Absolute Error (MAE) and
Root Mean SquareError (RMSE) [105], which are defined as follows:
M AE = 1 | ˆ R | (cid:88) ˆ r ui ∈ ˆ R | r ui − ˆ r ui | RM SE = (cid:118)(cid:117)(cid:117)(cid:116) | ˆ R | (cid:88) ˆ r ui ∈ ˆ R ( r ui − ˆ r ui ) (1)where r ui is the actual rating of user u for item i and ˆ r ui is the predicted rating. E2. Ranking accuracy assumes that items are ranked in a decreasing-rating order for each user and only the top itemsare presented. So, they evaluate the similarity in the order of rated items, providing a more robust evaluation methodthan MAE or RMSE. Pearson ( c ) and Spearman ( ρ ) correlation coefficients measure the linear relationship betweentwo (parametric/non-parametric) variables, as defined by the following equations: c ( x, y ) = 1 N − (cid:80) Ni =1 ( x i − ¯ x )( y i − ¯ y ) s x s y ρ ( u, v ) = 1 N − (cid:80) Ni =1 ( u i − ¯ u )( v i − ¯ v ) s u s v (2)where x i and y i are the i th elements in the variables of interest, and ¯ x and ¯ y are the sample means. Similarly, s x and s y represent the sample standard deviation for a sample of size N . u and v are the ranked variables counterpart of x and y . The values vary between -1 and 1 where the former exhibits a strong negative relation between two variables,and the latter being a positive one. A value of 0 indicates the absence of any linear relationship.Such metrics can evaluate the recommender system as a whole, regardless of the used rating scale [106]. They enabledevelopers to assess the recommender’s ability to provide ratings for energy suggestions that are consistent withexisting users’ ratings, thus, evaluating the quality of the given suggestions. According to an experimental work doneby [107], both Pearson and Spearman produced quite similar results, thus, being redundant if both used. Therefore,one can only use either in the context of recommender systems. E3. Information retrieval metrics , such as precision, recall, and F1 score, are among the most relevant metrics thatcan be used in recommendation systems context with slight modifications. Since the item ratings are usually in a 1-5scale, a threshold value T is used to convert an absolute rating to a binary prediction that classifies whether the item isrelevant or not [106].The metrics are depicted in equations (3-5). In the context of recommender systems, true positive (TP) indicates thatthe recommended energy saving action is relevant within the user context, or has been accepted by the user, while falsepositive (FP) means that the recommended action is irrelevant. On the other hand, false negative (FN) indicates thatthe recommender system failed to recommend an energy-efficient action, that was actually performed by the user, andlastly, true negative (TN) refers to a system that does not provide energy-efficient suggestions when the context doesnot demand one. Recall = T PT P + F N (3)
P recision = T PT P + F P (4)10
PREPRINT - F
EBRUARY
16, 2021 F Score = 2 × P recision × RecallP recision + Recall = 2
T P T P + F P + F N (5)To elaborate further, if an excessive energy consumption is detected, multiple recommendations can be relevant toreduce the energy consumption. For instance, if a television or a room’s lighting were left on, while no one is inthe room, the recommender system advising to turn-off either the television or the room’s lighting would be relevant(i.e. TP) to reduce household’s energy consumption. On the other hand, FP would be suggesting to turn-off either thetelevision or the lighting in a room where someone is currently within.
E4. Serendipity and freshness measure the novelty and variability of items recommended to the users in an effort toavoid recommending the same items again and again. Serendipity measures the amount of surprise an accepted actiongenerates for a user [103]. As if the recommender system is informing the user about a new piece of informationaccompanying an action, which is relevant but he/she has not heard about before. From this perspective, a distancemetric adapted from [103], defined by equation (6), is utilized to check the serendipitous virtue of the recommendersystem. It calculates the distance between a suggested action a and a set of previous suggestions A that the userconsidered. n A, c is the number of suggested actions within the same group c ∈ C in A , n A is the maximum numberof suggested actions from a single class c in A , and c ( a ) is the class of action a . d ( a, A ) = 1 + n A − n A, c ( a ) n A (6)Equation (6) estimates the distance a recommender system accumulates for a list of suggested actions. As an example,imagine a user i was subjected to a list of actions A throughout a certain period. By dividing this set to two subsetsof suggested observed actions A io and hidden ones A ih , the A io can be utilized by the recommender system to generateenergy-efficient action suggestions. Now, if the recommender is asked to generate 10 actions, we would like therecommender to suggest valid and relevant suggestions to the user i , which are NOT in the A io set, thus, increasingthe system’s serendipitous virtue. This is measured by calculating the distance score d ( a, A io ) for each a ∈ A ih , wherethe score will be reduced if actions from the same class as a , depicted in c ( a ) , are numerous, i.e. n A, c ( a ) is large. Itis worth noting that, in a sense, serendipity battles the accuracy of the recommender system, thus, it is important toperiodically check the relevance of suggested actions, as users can refrain from using the recommender system if itkept suggesting irrelevant ones [103].Freshness, on the other hand, indicates the recommender capability in suggesting new recommendations each timethe user interacts with [108]. However, since the given actions to reduce energy consumption are limited by naturein small households, it is possible to adjust the definition such that the same action is not recommended in a matterof few hours/days. Thus, equation (6) can be revisited and utilized every k hours/days to ensure that an action a wasminimally suggested in the previous k hours/days. E5. Acceptance ratio (AR) aims to quantify the agreement that a certain user exhibits to energy efficiency suggestionsprovided by the recommender system. In numbers, the ratio allows the recommender system to understand the under-lying probability of accepting the suggestions it provides, either on a holistic-level, or for a certain energy efficiencysuggestion, e.g. turning off the air conditioner, as seen in equation (7). AR = 1 C C (cid:88) c =1 AR c AR c = 1 M M (cid:88) i =1 a i,c r i,c (7)where r i,c is a recommendation and a i,c is either 0 or 1 depending on whether the recommendation r i,c is accepted. M is the number of recommendations per energy efficiency suggestion belonging to the same group C .The importance of this metric prevails when the recommender decides to send a suggestion, where it helps in answer-ing the question: Which energy saving suggestion should the recommender system send, an extreme energy savingsuggestion with low acceptance rate, or a moderate one with high acceptance rate? E6. Other metrics aim to evaluate: i) the
Coverage of recommendations in terms of the item or user space, whichmeans that the recommender systems must suggest all possible actions and recommend actions to all users, ii) the
Confidence of the system in its recommendations, iii) the
Trust of users to them, iv) the
Utility , v) the
Risk , vi) the
Robustness and many more. [36] 11
PREPRINT - F
EBRUARY
16, 2021
Aiming at increasing the acceptance ratio of energy saving recommendations, recommender systems should providethe user with incentives, to encourage and motivate them in promoting more sustainable energy behaviors. In thiscontext, several incentives have been deployed in energy efficiency recommendation systems, such as feedback rating,green currency and social comparison.
I1. Feedback rating is of significant importance, because it can show the end-users’ satisfaction regarding the de-livered recommendations [109, 110]. Therefore, gathering feedback ratings helps to efficiently adapt the recommen-dations to the users’ preferences and interests, and thereby results in a substantial contribution to deliver sustainablereductions in energy consumption.
I2. Green currency is tightly coupled with the rapid development and prevalent use of blockchain and cryptocurren-cies. The energy research community has also inspired the creation of green and/or CO coins, which can be usedas motivation towards energy efficiency. For instance, in [111], end-user engagement is promoted via the adoption ofmonetary rewards, named CO credits, in recognition of energy savings and effective achievements. I3. Social comparison has recently been adopted in various recommender systems with different applications (e.g.tourism and travels, movies, e-commerce, e-learning, healthcare, etc.). In the energy efficiency scenario, social com-parison aims to motivate electricity saving in households [112]. This research area relies on the use of fundamentaltheories of social psychology which can teach us on how to encourage end-users to preserve energy by recommendingindividuals with better profiles as a reference [113, 114]. More specifically, normalized comparison modules (thatperform consumer rankings) are incorporated in the eco-feedback tools, in order to help end-users to compare theirenergy usage patterns with those of their peers and neighbors. The effectiveness of this incentive mechanism comesfrom the fact that end-users are highly influenced by engagements and rankings of their peers on social networks[115, 116].All in all, Table 1 summarizes the characteristics of the recent and relevant energy efficiency recommender systems.It highlights their advantages, and outlines the main information about their taxonomy, which could be extracted fromthe overview conducted above.
First of all, based on the overview of existing energy efficiency recommender systems conducted above and by ana-lyzing the summary in Table 1, it has been clearly demonstrated that most of the frameworks have been implementedusing cloud computing. This is due to its different advantages, among them is its flexible lease and release of comput-ing resources as per the end-user’s requirement [121]. However, cloud computing has other issues, such as the privacypreservation problem, which could be occurred whenever the data may outbreak the service provider and the informa-tion could be deleted purposely [122]. Additionally, technical issues could also occur due to the fact that servers couldbe down, and hence, it becomes hard to gain back access to the needed resources/data at the right moment and fromanywhere. For instance, non-availability of services could be a result of a denial-of-service attack (DoS) [123].Furthermore, it is worth noting that most of the recommender systems have been used to polish energy consumptionbehaviors in households while there are also other frameworks that discussed their utilization in commercial buildings,work spaces and other types of public buildings (e.g. hotels and hospitals). While for the objectives of the recom-mender systems, almost the same attention has been given for developing both action recommendations and strategyrecommendations in existing energy saving recommendation systems frameworks.On the other hand, recommender systems for energy efficiency face other difficulties and issues, which need to beovercome while developing reliable solutions. In this section, we focus on introducing them along with discussingthe commercialization potential of energy saving recommender systems and related issues, i.e. identifying the mainmarket barriers and market drivers [124].
Although there has been a significant progress in developing recommender systems as discussed above, various is-sues that hinder the establishment of effective recommendation engines still exist. The most critical problems anddifficulties that exercise a negative impact can be summarized as portrayed in Fig. 4. Explicitly, it outlines the maindifficulties and commercialization issues discussed in the context of energy saving recommender systems.12
PREPRINT - F
EBRUARY
16, 2021Table 1: Summary of the relevant energy efficiency recommender systems proposed in the literature.
Framework Type of RS Advantages Computing Objective ApplicationPlatform EnvironmentPinto et al. [40] M1 Recommend operational light intensity P1 O1 Householdssatisfying users’ comfortKaur et al. [41] M1 Predict energy consumption ratings and P2 O2 Householdsoffer personalized recommendationsSchweizer et al. [47] M1 Learn the energy usage habits and P1 O1 Householdsinterests of consumersZhang et al. [52] M2 Generate tailored recommendations P1 O2 Householdsfollowing predicted ratings of energy usageZheng et al. [53] M2 Provide appliance-level consumption P3 O2 HouseholdsrecommendationsReViCEE [46] M2 Predict collaborative recommendations of P4 O1 Householdslight preferences of end-usersGarcia et al. [117] M2 Produce tailored advice on end-users’ P1 O2 Householdsactivities similar scenariosShigeyoshi et al. [58] M3 Produce contextual based advice with P3 O2 Householdssocial experiment ratingsLuo et al. [118] M3 Tailored recommendations with textual P1 O2 Householdsappliance advertisementsWei et al. [44] M3 Provide move and shift-schedule P1 O1 Commercialrecommendations buildingsREHAB-C [119] M3 Tailored recommendations with feedback P2, P3 O2 Academicon end-users’ preferences buildingsStarke et al. [60] M4 Provide Rasch profile based recommen- P3 O1 Householdsdations of end-users’ behaviorStarke et al. [61] M4 Generate Rasch profile recommendations P3 O1 Householdsbased on a social experimentLi et al. [64] M5 Provide tailored recommendations to support P1 O1 Work spacesthe use of renewable energy solutionsWei at al. [45] M5 Optimize power consumption using P1 O2 Commercialhuman-in-the loop model buildingsBravo et al. [120] M6 Create energy saving recommendations P2 O1 Householdsbased on electricity priceAiello et al. [74] M6 Fuse contextual information, mathematical P2 O1 Public buildingsformulation and experts’ knowledgeKiran et al. [78] M7 Alleviate the cold start issue P1 O2 HouseholdsPinto et al. [79] M7 Improve predictions of consumers’ P3 O2 Public buildingspreferences and recommendationsRocha et al. [83] M8 Decision aggregation of heating and cooling - O2 Public buildingssystems operations with energy demandAnvari et al. [84] M8 Multi-agent based optimization P1 O2 Households/public buildings
Recommender systems are mainly based on the analysis of historic consumer data, which usually comprise few cus-tomer demographics and mainly customer ratings for items (or actions). Because a consumer may only rate a smallnumber of the actions that are available on the recommendation platform, this leads to a sparseness on the ratings forsome actions or users [125, 126]. Explicitly, this results in producing unreliable recommendations, which in turn couldreduce consumer satisfaction [127].
This issue refers to the lack of data (i.e. ratings) for new consumers or new actions. The problem is more acute in thestart of the recommendations generation process, when the platform still has a few consumers and limited information13
PREPRINT - F
EBRUARY
16, 2021Figure 4: A summary of limitations and difficulties of energy saving recommender systems.about their preferences [128]. Explicitly stated user preferences can be used to enhance customer profile, and usersimilarity metrics can be used to assign newcomers to existing customer clusters with similar preferences. Anotherproblem occurs in systems that refresh their catalog of items or actions. When a new action is added as an option tothe system, it does not have any ratings yet, so collaborative methods cannot recommend them. Therefore, this leadsto the fact that an action cannot be recommended easily, and hence, less likely to be noticed by consumers [129, 130].Probabilistic techniques that recommend new items with lower probabilities and rule-based recommenders can be usedin place, to tackle the cold start problem.
Due to the large variability in the type of buildings and the respective objectives for energy savings, when developinga recommender system for this purpose, it is of utmost importance to use energy consumption datasets form differentenvironments (e.g. households, commercial buildings and industrial areas) to evaluate the developed systems andefficiently generate personalized recommendations. In this regard, datasets are utilized as benchmarks for developingnew recommender models and comparing them to existing systems under the same conditions. Therefore, datasets playa major role in the creation of successful recommendation systems, and most of the effective and robust recommendersystems are those built upon large-scale datasets including big amounts of consumers’ data [131]. However, forthe energy sector, another important issue that impedes the generation of efficient recommendations is the absenceof appropriate datasets along with the difficulties encountered to collect them. Moreover, the lack of open-accessrepositories containing existing datasets make the recommender systems comparison very difficult, or even impossible[132, 133].
As presented in Section 3.2, various recommender systems have been developed and utilized for generating energysaving recommendations. However, the comparison of their efficacy is a daunting task since the assessment resultscould hardly be reproducible due to the absence of toolkits that support such tasks [134]. Explicitly, the actual issuesthat hinder reproducing recommendation systems results had put recommender system community in a problematicsituation [135, 136]. Researchers and developers (who require effective recommender algorithms and baselines againstwhich to compare novel approaches) usually obtain very limited guidance in existing research and development sources[137]. In this line, to alleviate these issues and enable reproducibility, recommendation systems community needs to(i) review other research topics and inspire from them; (ii) establish a common definition of reproducibility; (iii)capture and comprehend the decisive factors that impact reproducibility; (iv) perform more extensive experiments; (v)promote developing and using recommender frameworks; and (vi) launch experimental platforms that includes recentstate-of-the-art algorithms and datasets and establish best-practice guidelines for recommendation system research,which will also ensure a fair comparison among systems [138, 139].14
PREPRINT - F
EBRUARY
16, 2021
Energy conservation and energy security have been big concerns in recent years. Energy deficiency does not only affectthe economy, culture and development of the world, but also results in global warming. This is why recommendersystems play an imperative role in transforming end-user behavior towards improved energy efficiency. Therefore, it isof fundamental importance to investigate the main market drivers that will force the incorporation of energy-efficientand cost-effective technology, and to perform activities that help in controlling the decision-making process for energyefficient buildings, thus, providing an incentive for energy conservation [140]. An important driver is the EU energyand climate package [141] for the year 2030 that includes the goals of 40% reduction of greenhouse gas emissions(compared to 1990 levels), and 27% increase of renewable energy sources (RES) in the EU-27 energy mix (today6.5%) and 27% mitigation in the primary energy consumed (saving 13% compared to 2006 scales). These goals haveled the inception of many energy efficiency research initiatives, some of which incorporate recommender systems [68].A global driver, that is derived from the same incentives, is the strategic decision, made by the top management ofseveral firms, to comply with the new energy saving plans. Such decisions include energy reductions within the firmitself, optimizations of the products’ and/or services’ production pipeline, and efforts to reduce the overall carbonfootprint. The TARHSEED initiative, is a bright example: the governmental initiative’s aim to raise awareness onenergy saving activities and reduce unnecessary energy usage in the country as a whole [142, 143], has been boostedby several advertising initiatives, standards and competitions, which motivated both individuals and corporations tooptimize their energy efficiency.
The technology readiness level (TRL), the compatibility (i.e. how feasible it is to integrate it with existing systems),and the business model behind a technological solution are crucial factors in its ability to change its target market.The same factors hold for energy saving recommender systems and affect their impact to the energy market landscape.The global energy landscape is changing, driven by the need to reduce emissions and increase the security of supplywhile increasing the intermittent renewable energy in the energy mix. In this new landscape, the increasing powerconsumption requires maintaining the power grid reliability, regulating electricity flow with less mismatching betweenelectricity generation and demand and reducing the energy footprint. The optimization of scheduling, the improvementof energy quality and assets efficiency, the integration of dynamic pricing and the incorporation of more renewableelectricity sources are among the continuous challenges of the traditional energy grid.For a recommender system to penetrate the market and establish its presence, there is a number of market barriers toconsider. First, technical barriers present a major milestone in adapting recommender systems in mainstream productsand services. Namely, there is a discrepancy between the different evaluation standards of recommender systems, i.e.there is no unified standard for objective benchmarking. Moreover, the lack of comprehensive datasets impedes theprogress in creating powerfully dynamic recommenders.From a legal standpoint, several market barriers exist, which in the context of energy efficiency, comprise a numberof regional standards and regulations for residential and industrial energy consumption. For example, in Greece, theTransposition of the European Directive 2009/28/EC was established in 2010. In order to acquire a new permit tobuild in 2011, it is either appropriate to obtain an annual solar percentage of 60% for the development of hygienic hotwater from solar thermal systems or to explain technological challenges in the event of non-compliance. This, in turn,puts additional obstacles towards commercializing recommender systems as they have to comply with many standardsto obtain international recognition.
The focus in this section is to provide insights on where the actual energy recommender systems research is headingto as well as the related challenges attracting considerable R&D in the foreseeable and far future. Specifically, Fig. 5summarizes the current challenges and future orientations of energy efficiency recommender systems, which havebeen addressed in this framework. 15
PREPRINT - F
EBRUARY
16, 2021Figure 5: Currant challenges and future orientations.
The recent success of machine learning (ML) techniques in solving daily prediction or classification tasks has dra-matically influenced the number of applications that adopt ML models, using them as black boxes that makes themdifficult to understand by the end-users [144]. The ability of an ML model to “explain itself and its actions” to theusers is considered to be an emerging important factor for the current modern AI applications transitioning to modernexplainable AI models [145, 146], as depicted in Fig. 6.Figure 6: The transition to explainable AI.In general, there are five key concepts that can describe each recommendation task and are referred to as “the five Ws of recommendations”: W hen, W here, W ho, W hat and W hy [147]. The five Ws usually represent the time-drivenrecommendations ( when ), location-driven recommendations ( where ), their social component ( who ), application-drivenrecommendations ( what ), and their explanations ( why ), respectively.Following the emergence of explainable AI, the goal of “Explainable Recommendation Systems” is to offer helpfulsuggestions to consumers, accompanied by explanations that generally relate to the reasons for providing these recom-mendations or the advantages of selecting the suggested alternatives [148, 149]. The key contribution of these reasonsis that they will boost the system’s persuasiveness, customer comprehension and happiness and offer an instant rewardto the user.Latest work on explanation-driven recommendations is centered around two core aspects: 1) the type of explanationsgenerated (e.g. textual, visual, etc.); and 2) the employed algorithm or model to generate the explanation, which canbe loosely classified as matrix factorization, subject modeling, graph-driven, deep learning, knowledge-graph, inter-action rules, and post-hoc models, etc. [147]. Explainable recommendations, classified by the nature of explanationproduced, are enumerated as follows:• User-based and item-based explanations : This is a traditional type of explanation based on user’s feedbackand is expressed as a statement of similarity among the system’s different users (in the case of user-based) oritems (in the case of item-based recommendations).•
Content-based explanation : This type is solely based on the item’s feature space (e.g. for book recommen-dations the book type, the writer, etc.). 16
PREPRINT - F
EBRUARY
16, 2021•
Textual explanations : In the textual explanations, the recommendations include explanation sentences thatmay be based on other users’ reviews or natural language processing techniques.•
Visual explanations : Visual explanations utilize item images for explainable recommendations indicatingthe part of the image that matches the item images that the user might be interested in.•
Social explanations : These explanations refer to items that user’s “friends” in the social networks or in aspecific community also prefer.•
Hybrid explanations : Hybrid explanations refer to combinations of one or more of the previous types ofexplanations.Focusing on energy saving recommender systems, currently we identify the lack of recommendation systems in thearea of energy sustainability, which adopt the recent trends of explainable AI. Recent surveys like [150] try to overviewexisting methods to set the current research issues related to the explainable recommendations. For instance, [151]proposes a deep explicit attentive multi-view learning architecture for modeling multi-level characteristics of expla-nations, or the framework in [152] that examines another scheme to create a set-based recommendation platform togenerate textual and transparent explanations of film recommendations. Aiming at developing a knowledge-basedscheme to create explainable item recommendations, a technique to leverage external knowledge is proposed in [153],which is based on adopting knowledge graphs when information from content and product/item reviews is unavailablefor generating explanations. Interpretable models, are based on transparent processes to decide the recommendationlists, hence, it is easier for generating appropriate explicit feature-level explanations to justify the reasons behind therecommender’s suggestion for particular items [154]. Following the same scenario of graph-based models, He etal. in [155] introduce a technique that could rank the vertices of a tripartite graph and furnish explanations for thetop-ranked, aspects-target, and user-recommended item triplets. By contrast, in the field of energy saving and recom-mendations for energy-related behavior, there is limited literature that elaborates on the rules of producing a particularrecommendation. Authors in [156] propose a user-centric and visual analytics approach for developing an interactiveand explainable forecasting and analysis of electric power demand in prosumer settings. Moreover, it has been ad-vocated that this would be endorsed by behavioral analysis to enable the treatment of possible relationships betweenenergy usage footprints and the interaction of prosumers with energy analysis tools, including customer portals andrecommendation systems.
Most of the actual tailored energy recommender systems are, by and large, limited in terms of effectiveness [157].To improve their performance, the incorporation of cognitive and behavioral knowledge, is essential, where tailoredrecommendation frameworks can be friendlier and more human-centric [158, 159]. Thus, this results in eventuallyenhancing users’ experience and loyalty and increase their satisfaction. To that end, more effort should be establishedin this direction aiming at developing psycho-cognitive method recommendation systems that generate personalizedenergy saving actions and advice based on consumers’ preferences, emotional states and centers of interest [160, 161].Accordingly, psycho-cognitive recommender systems are new intelligent recommender systems that help in (i) com-prehending end-users’ preferences; (ii) detecting changes in end-users’ habits and attitudes through time; (iii) pre-dicting end-users’ unknown choices and behavioral change; and (iv) investigating adaptive techniques for generatingintelligent recommendations within a changing environment [162, 163]. All these tasks mixed together could improveend-users’ behaviors and increase their awareness towards a more sustainable and energy-efficient usage [164, 165].A typical representation of psycho-cognitive recommender system is illustrated in Fig. 7, in which three importantmechanisms called knowledge-driven, cognition-driven and data-driven are used to develop a psycho-cognitive rec-ommendation framework.The same category also includes recommender systems that build on persuasiveness in order to maximize acceptance[166]. In some cases, the messages sent to the user are also personalized to match user’s preferences and values [24].Active learning is a key component in such approaches because it allows the system to continuously adapt to thechanging user needs and demands [167].
The preservation of user privacy is a key requirement in several recommender systems, especially in online socialcommunities. Several techniques have been proposed in the past, ranging from k-anonymity, to differential privacyand homomorphic encryption. This kind of frameworks could be split into three main groups: (i) perturbation-basedtechniques that introduce noise to the existing data [168], without affecting the final recommendation result; (ii)encryption-based schemes that transform the original information within the main recommendation technique (e.g.17
PREPRINT - F
EBRUARY
16, 2021Figure 7: Typical representation of a cognitive recommender system for energy saving.within the matrix factorization component [169]); and (iii) techniques that develop novel matrix factorization al-gorithms under local differential privacy (LDP) [170]. In the case of content recommender systems, group-basedapproaches [171] implement the principles of k-anonymity in order to maintain recommendation efficiency withoutaffecting user privacy. Similar challenges are set for navigation solutions that rely on crowd-sourced data collection[172].Another essential challenge in privacy-preservation is servers-related, and tackles their features, especially when theyare unreliable (untrusted) or comprise security weaknesses (vulnerabilities), and thereby collecting consumers’ feed-back may result in cyber liability owing to data leakage [173]. Early works in the privacy of consumer data in electricload monitoring applications [174] mostly focus on non-intrusive monitoring techniques to combat potential invasionsof privacy [175]. Later works minimize the amount of collected reference data through sampling. For example, in[176] authors remove redundant energy traces, which do not contribute new knowledge to the recommender system.To the same direction end, privacy-preserving recommendation approaches aim at preserving consumers’ privacythrough hiding their rating feedback from servers and/or other consumers [177, 178]. Fig. 8 presents a typical rep-resentation of an energy saving recommender systems. It illustrates what are the sensitive information that need tobe encrypted before submitting them to the recommender system’s server, such as energy consumption data (providesthe intruder with information about the presence/absence of the end-user in his household), end-users’ feedback andratings and private information (personal data, location, number of end-users, etc.) [179, 180]. After storing andprocessing collected data, the recommender system encrypts the generated recommendations before sending them tothe targeted end-user. Moreover, end-users could collaborate with each other to compute action similarity using theirprivate keys. On the other hand, with the arrival of the blockchain technology, new opportunities have been opened upto develop a novel generation of recommender systems that can overcome the privacy-preservation issues and protectconsumers’ data [181, 182]. For example, Bosri et al. propose Private-Rec, which is privacy-preserving recommendersystem using AI and blockchain [183]. Explicitly, blockchain has been used to provide the end-user with a securemechanism using the distributed attribute where data could be exchanged with the required permission. While in[184], blockchain is deployed as the backbone of a decentralized recommender system, in which a secure architecturehas been introduced using decentralized locality sensitive-hashing classification along with recommendation genera-tion.
The concepts of time-aware and concept-aware recommendations have been widely discussed in the recommendersystems’ community. From the early works on movie recommendations [185], to the more recent works on time-aware point-of-interest recommendations [186, 187], several frameworks for modeling, computing and presentingtime-aware recommendations have been proposed in the general domain [188, 189].Recommender systems for the energy sector differ highly from those used in other research topics. Explicitly, mostexisting models concentrate on recommending energy saving actions that fit consumers’ preferences while putting a18
PREPRINT - F
EBRUARY
16, 2021Figure 8: A typical privacy-preserving recommender system based on encryption and collaborative filtering.slight importance on the temporal information and its influence on the recommendations [190]. In this respect, we as-sert that further attention should be paid for time-aware recommendations in energy saving applications in buildings topush them into the foreground [191, 192]. This kind of recommendations is more appropriate to emergency situations,e.g. the case of the Coronavirus COVID-19 pandemic, where real-time and time-aware recommendations should beprovided according to the current situation. Explicitly, due to the mass restrictions imposed on people’s movement, therise of teleworking and online learning has led to an increase of energy consumption in domestic buildings [165, 193].On the other side, with the widespread use of ML tools, using deep learning models would be a promising approach todevelop recommender systems that encompass contextual information into neural collaborative filtering recommenderframeworks [194, 195].
As we are in the big data era, modern energy saving recommendation systems face tremendously increasing datavolume and complexity due to the use of a massive number of connected devices. Traditional computation algorithmsand experiences on small datasets may not be efficient today. Therefore, developing robust recommender systemsthat is capable of processing large-scale data, is becoming a challenging endeavour for several applications. Authorsin [38] provide an interesting survey on the challenges and solutions for recommender systems for large-scale socialnetworks. Big data, variety, and volume are the three major challenges for recommender systems in large socialnetworks, which bring state-of-the-art collaborative filtering algorithms to their limits. Additionally, the large volatilityof social network data (e.g. new users and items added or removed on a daily basis) has raised the interest for newevaluation metrics, that promote recent [196] and diverse [197] entries (e.g. diversity, freshness, serendipity, etc.) andtackle the cold start problem.From the ML and deep learning perspective [198], graph convolutional methods are gaining the researchers’ interest[199], since they can summarize the graph structure of social networks and combine it with the lateral informationthat may be hidden in the items or in the relations among them. Compact latent factor models that combine contentwith ratings [200] prove to be more efficient than simple collaborative filtering algorithms in tackling the cold startproblem. In order to balance the exploration-exploitation dilemma (exploit interesting items while exploring newitems) and continuously capture user feedback without relying on item context, multi-armed bandit approaches [201,202, 203, 204] have been recently proposed. 19
PREPRINT - F
EBRUARY
16, 2021In order to solve the technical issues that may arise from the scalability of recommender systems in large datasets,several parallel and distributed algorithms have been proposed, which either rely on the splitting of the dataset, usingsocial or other information [205, 206] and its parallel processing, or on the refactoring of existing algorithms inorder to take advantage of the use of graphical processing units (GPUs) [207, 208, 209]. The issue of big datahandling has also been studied in the domain of energy efficient recommender systems for recommending energy plans[210], providing actionable recommendations [157] or improving comfort and energy efficiency in tandem [211]. Thesolutions discussed so far in the pertinent literature focus on data sampling or compression.
This section highlights the most promising research directions that will have a significant impact on improving theeffectiveness of energy saving recommender systems in the near future.
As described before, increasing the user’s trust and transparency to the system is an important concept in modern MLmodels and also a tool for maximizing the recommendations’ acceptance in modern recommender systems. Thus,in the field of recommender systems for energy saving, the system has to accompany every recommendation for anenergy saving action with:1. an explanation of why this particular advice is suggested2. a statistical fact on what are the benefits following the recommended actionHence, we have introduced the (EM) explainable recommender system for energy efficiency , and described the twomost essential characteristics that recommendation explainability is based on, as depicted in Fig. 9 [119, 212]:1. Reasoning:
This feature explores the context of global recommendations and seeks to provide thoroughjustification of why each recommendation has been made. It may be metadata of the end-user context (e.g.the end-user is not present in the room), about the device consumption (e.g. it is turned on for a long time) orabout the external circumstances that cause the appliance to be switched off.2.
Persuasion:
This aspect draws on end-users’ expectations, motivations and long-term values, and adoptstheir ratings to pick the most suitable and relevant explanation about each recommended intervention.Based on this strategy, and adopting a hybrid type of explanation in our approach, we enhance a persuasion fact alongwith the textual explanation in the recommendations’ body. Using this approach, we try to provide the actual benefitthe recommended action will achieve for the end-user, in an attempt to persuade him/her to increase his acceptanceover the provided recommendations. Initial results of our evaluations show that such an approach can impact users’trust to the system and can bring an increase of 20% in the acceptance ratio of provided recommendations.
In this technological age, it is not a secret that humans are attracted to imagery type of media, much more than the onesof textual nature. This can be witnessed by how millennials are exploiting technology nowadays. With this in mind,it can be argued that in order to have a better recommendation dialogue with end-users, aiding such recommendationswith visualized charts and evidence can significantly aid in making them persuasive. By stating this, it is by no meansindicating that the textual recommendations, i.e. explanations provided by the recommenders, should be discarded, butrather they are complementary to one another. Visualization and textual interpretations hand-in-hand can be integratedfully to structure suggestions given by the recommender systems, which are deemed to influence behavioral change.All the following discussed frameworks have used visualizations one way or the other to influence behavioral changein their system.A semantic smart home system for energy efficiency, namely SESAME, is proposed in [213] to provide daily andmonthly overlays of energy consumption data, CO footprint and financial impact. Also, the user interface (UI) tocontrol the appliances and create certain rules is also shown. On the other hand, Fern´andez et al. [214] showcase aheatmap exhibiting, in hourly fashion, the usage of air conditioning facilities for a whole week. Similarly, friendly UIis also provided to allow users more control over given services through smart devices. On the other hand, enCOM-PASS framework push visualizations aided by context-aware collaborative recommender systems on mobile platforms (EM) : Consumer Engagement Towards Energy Saving Behavior by means of Exploiting Micro Moments and Mobile Rec-ommendation Systems ( http://em3.qu.edu.qa/ ) PREPRINT - F
EBRUARY
16, 2021Figure 9: The flow of explainable recommendations generated in the (EM) framework.to provide energy-efficient recommendations from socio-technical point-of-view [215, 216]. Moreover, the frameworkalso created two games on smartphones to teach children about the importance of rationalizing the consumption ofboth water and energy. The former being taught through “Drop! The Question” application, which is developed withSmartH2O framework, and the latter through “Funergy” application [217]. Entropy, another framework, provides con-ventional time-series visualization for sensor data streaming in a desktop application to aid the recommender system[218], while both Bernard and HEMS-IoT create a mobile application for that purpose [219, 220]. CASER frameworkproduces both web and mobile applications for data visualization but showcase variant visualization including time-series and heatmap charts on both household and substation levels (multiple households) [221]. Lastly, (EM) createstwo distinct applications, where the first application on iOS showcases data visualization in recommendations [222]and the latter on both Android and iOS, developed on React Native, studies the effect of different charts on end-usersunderstanding [223]. Fig. 10 depicts the flowchart of the visualization recommender system developed in the (EM) .From the previous illustrated work, three important prospects can be further investigated when integrating the datavisualization pillar with the recommendations. Firstly, further studies can be established to understand the impact dif-ferent visualizations have on end-users as in [224]. Not only that, but also create novel data visualizations specificallyfor energy consumption data, which are deemed simple for end-users from different backgrounds. Secondly, using thevisualization graph hand-in-hand when the recommendation is suggested. In other words, the recommender systemrefers to the visualization and demonstrates the anomalous consumption through such visualizations. Thirdly, with therecent pandemic humanity is facing (i.e. COVID-19), people are working from home for social distancing, and thus,the energy consumption has increased in the domestic sector [225]. Therefore, it would be more important than everto generate personalized and timely recommendations. Non-intrusive load monitoring (NILM) has been deployed in different energy saving projects instead of submeteringfor detecting appliance-level consumption data and other related information, e.g. when exactly a specific applianceis turning on/off without the need to install further submeters [226, 227]. In this line, collecting energy consumptionfeedback using NILM has turned around good performance at low/no cost. On one hand, developing efficient energysaving recommender systems relies on an accurate analysis of energy consumption data, especially at the appliance-level, and on the detection of abnormal energy usage. Tailored recommendations are then produced following thefeedback of the anomaly detection stage. Thus, the development of non-intrusive recommendation systems usingsubmetering is a promising solution to be investigated, which is not scalable because it needs to analyze individual21
PREPRINT - F
EBRUARY
16, 2021Figure 10: Example of the visual recommender system proposed in the (EM) framework.appliance consumption traces [228]. On the other hand, for energy data analysis and visualization, NILM has beenconsidered as a scalable and practical alternative to submetering. However, the use of NILM in recommender systemshas not been discussed before since the aim of NILM is to provide appliance-level energy footprints. In this regard, itis of significant interest to assess the signal fidelity of devices’ fingerprints generated by existing NILM algorithms todevelop effective non-intrusive recommender systems. This could figure out end-users’ preferences and related infor-mation as well [229]. Consequently, by using NILM instead of submetering, the development cost of recommendersystems will significantly be reduced [118, 230]. Energy recommender systems have become an essential solution for energy efficiency in buildings. While a largenumber of existing frameworks are focused on using cloud-to-edge architectures, in which recommended energy sav-ing actions are transmitted to the edge device (e.g. consumer’s smartphone) after completing the computing task inthe cloud server [231]. Although these architectures allow to achieve a good efficiency, they are prone to seriousnoticeable delays in the system’s feedback and user’s reaction because of the network bandwidth and latency betweenthe cloud and edge [232]. By contrast, implementing the recommender algorithms directly on the edge can allowreal-time computing and identify consumers’ interests/preferences more accurately and thereby increasing their satis-faction and trust of the generated recommendations [233, 234]. Thus, recently, a great deal of attention is devoted todevelop and implement recommender systems on the edge and/or on fog devices, which can tremendously reduce thecomputational time, minimize the cost of cloud hosting and ensure privacy-preservation [235, 236, 237].To that end, various frameworks have been proposed in different research fields to investigate the applicability ofedge-based and fog-based recommender systems [238, 239]. For instance, in [100], aiming to satisfy the new re-quirements of recommender systems, e.g. the low latency and uninterrupted service, Wang et al. propose a fog-basedrecommendation framework based on collaborative filtering. It can overcome the problem of information overloadin fog computing and help in generating personalized recommendations for improving system performance. In thesame manner, a fog-based recommender system which helps to bridge the gap between the cloud and end-devices isproposed in [240]. This system has been used to improve the performance of the E-Learning environments. Further-22
PREPRINT - F
EBRUARY
16, 2021Figure 11: Example of a typical representation of a fog-based energy recommender system.more, in [241], a fog-based recommender system is introduced to provide recommendations regarding the lifestyle,dietary plans and exercises to an ensemble of cardiovascular disease patients. While in [231], an edge-based recom-mender system is proposed to allow (i) capturing real-time end-users’ preferences more precisely; and (ii) generatingpersonalized and satisfying recommendations. Fig. 11 describes a typical representation of a fog-based recommendersystem for multiple users, in which a hybrid computing scheme, using cloud and fog servers, is generally adopted toimplement the main tasks of the recommendation framework.
This article presents a critical review of recommender systems for energy efficiency in buildings. Accordingly, ataxonomy of recommender systems is firstly conducted based on different aspects. Following, a critical analysis isconducted to highlight their strengths and limitations before deriving the current challenges and cutting-edge topics,which can be targeted in the near future to improve their performances.By and large, energy saving recommendation systems are proving to be a promising solution to promote sustainabilityand reduce carbon emissions, especially with the widespread deployment of smart-meters, IoT sensors and ML tools.Their evolution is accompanying the evolution of the intelligent Internet systems. The first generation of recommenda-tion frameworks were based on collaborative filtering, case-based, PRM and Rasch-based engines through analyzingonly energy-based data. While the second generation relies on the use of information fusion and deep learning mod-els, in which other kinds of data are gathered and transmitted to the recommender engine to be analyzed together withenergy consumption footprints. This movement, that the second generation is promoting for, helps in generating moreaccurate recommendations. 23
PREPRINT - F
EBRUARY
16, 2021Moving forward, we have also discussed the third generation of recommender systems for energy efficiency, whichrelies on adding other innovative modules into the recommender engines, i.e. explanations, visualizations and time-aware information processing. The use of edge computing technologies and edge AI is playing a major role in makingdevelopment real-time recommendation systems a reality. Moreover, this results in improving the quality and ac-ceptance of recommendations and increasing the end-users’ satisfaction. In this line, future research will focus onfostering the existing systems and technologies for improving both the quality and applicability of recommendationframeworks. Concurrently, novel directions of research will be furthered to develop a novel generation of highlyautomated recommender system via the use of (i) NILM strategies instead of conventional smart-metering; (ii) edgecomputing as an alternative to cloud computing; and (iii) privacy-preservation recommendation systems to increaseend-users’ trust.Finally, it is worth noting that the application of recommender systems in the building energy sector is a very promisingfield since it does not only recommend energy saving actions but can also be extended to help consumers acquireappliances. In this regards, several factors could impact the choice of the consumer, such as the energy consumptionof the appliance, its manufacturer, its price and other specifications. However, this will make the development of amore comprehensive energy efficiency recommender systems more challenging. In the grand scheme, recommendersystem will remain a strong pillar in the future of artificial intelligence and behavior change.
Acknowledgements
This paper was made possible by National Priorities Research Program (NPRP) grant No. 10-0130-170288 from theQatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsi-bility of the authors.
References [1] X. Cao, X. Dai, J. Liu, Building energy-consumption status worldwide and the state-of-the-art technologies forzero-energy buildings during the past decade, Energy and buildings 128 (2016) 198–213.[2] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Robust event-based non-intrusive appliance recognition usingmulti-scale wavelet packet tree and ensemble bagging tree, Applied Energy 267 (2020) 114877.[3] M. Economidou, V. Todeschi, P. Bertoldi, D. Agostino, P. Zangheri, L. Castellazzi, Review of 50 years of EUenergy efficiency policies for buildings, Energy and Buildings (2020) 110322.[4] P. Pylsy, K. Lylykangas, J. Kurnitski, Buildings’ energy efficiency measures effect on co2 emissions in com-bined heating, cooling and electricity production, Renewable and Sustainable Energy Reviews 134 (2020)110299.[5] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Efficient multi-descriptor fusion for non-intrusive appliancerecognition, in: 2020 IEEE International Symposium on Circuits and Systems (ISCAS), IEEE, 2020, pp. 1–5.[6] K. H. Refat, R. N. Sajjad, Prospect of achieving net-zero energy building with semi-transparent photovoltaics:A device to system level perspective, Applied Energy 279 (2020) 115790.[7] P. Huang, G. Huang, Y. Sun, Uncertainty-based life-cycle analysis of near-zero energy buildings for perfor-mance improvements, Applied Energy 213 (2018) 486–498.[8] Y. Lin, S. Zhong, W. Yang, X. Hao, C.-Q. Li, Towards zero-energy buildings in china: A systematic literaturereview, Journal of Cleaner Production (2020) 123297.[9] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, Recent trends of smart non-intrusive load monitoring in build-ings: A review, open challenges and future directions, International Journal of Intelligent Systems (2021) 1–28.[10] Z. M. Azizi, N. S. M. Azizi, N. Z. Abidin, S. Mannakkara, Making sense of energy-saving behaviour: Atheoretical framework on strategies for behaviour change intervention, Procedia Computer Science 158 (2019)725–734.[11] S. C. Staddon, C. Cycil, M. Goulden, C. Leygue, A. Spence, Intervening to change behaviour and save energyin the workplace: A systematic review of available evidence, Energy Research & Social Science 17 (2016)30–51.[12] A. Belhadi, Y. Djenouri, G. Srivastava, D. Djenouri, J. C.-W. Lin, G. Fortino, Deep learning for pedestriancollective behavior analysis in smart cities: A model of group trajectory outlier detection, Information Fusion65 (2021) 13–20. 24
PREPRINT - F
EBRUARY
16, 2021[13] P. Fraternali, S. Herrera, J. Novak, M. Melenhorst, D. Tzovaras, S. Krinidis, A. E. Rizzoli, C. Rottondi, F. Cel-lina, encompass—an integrative approach to behavioural change for energy saving, in: 2017 Global Internet ofThings Summit (GIoTS), IEEE, 2017, pp. 1–6.[14] M. Casals, M. Gangolells, M. Macarulla, A. Fuertes, V. Vimont, L. M. Pinho, A serious game enhancing socialtenants’ behavioral change towards energy efficiency. in 2017 global internet of things summit (giots) (2017).[15] C. Sardianos, C. Chronis, I. Varlamis, G. Dimitrakopoulos, Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira,Real-time personalised energy saving recommendations, in: The 16th IEEE International Conference on GreenComputing and Communications (GreenCom), 2020, pp. 1–6.[16] W.-S. Hwang, H.-J. Lee, S.-W. Kim, Y. Won, M.-s. Lee, Efficient recommendation methods using categoryexperts for a large dataset, Information Fusion 28 (2016) 75–82.[17] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Improving in-home appliance identification using fuzzy-neighbors-preserving analysis based qr-decomposition, in: International Congress on Information and Com-munication Technology, Springer, 2020, pp. 303–311.[18] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, C. Sardianos, I. Varlamis, G. Dimitrakopoulos, On the appli-cability of 2d local binary patterns for identifying electrical appliances in non-intrusive load monitoring, in:Proceedings of SAI Intelligent Systems Conference, Springer, 2020, pp. 188–205.[19] C. Becchio, M. Bertoncini, A. Boggio, M. Bottero, S. P. Corgnati, F. Dell’Anna, The impact of users’ lifestylein zero-energy and emission buildings: An application of cost-benefit analysis, in: International Symposium onNew Metropolitan Perspectives, Springer, 2018, pp. 123–131.[20] M. Ashouri, F. Haghighat, B. C. Fung, A. Lazrak, H. Yoshino, Development of building energy saving advisory:A data mining approach, Energy and Buildings 172 (2018) 139–151.[21] U. S¸ ims¸ek, A. Fensel, A. Zafeiropoulos, E. Fotopoulou, P. Liapis, T. Bouras, F. T. Saenz, A. F. S. G´omez, A se-mantic approach towards implementing energy efficient lifestyles through behavioural change, in: Proceedingsof the 12th International Conference on Semantic Systems, 2016, pp. 173–176.[22] I. Varlamis, C. Sardianos, G. Dimitrakopoulos, A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, Reshapingconsumption habits by exploiting energy-related micro-moment recommendations: A case study, in: Commu-nications in Computer and Information Science, Springer International Publishing, Cham, 2020, pp. 1–22.[23] O. Iweka, S. Liu, A. Shukla, D. Yan, Energy and behaviour at home: a review of intervention methods andpractices, Energy Research & Social Science 57 (2019) 101238.[24] C. Sardianos, I. Varlamis, C. Chronis, G. Dimitrakopoulos, A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, Theemergence of explainability of intelligent systems: Delivering explainable and personalized recommendationsfor energy efficiency, International Journal of Intelligent Systems 36 (2) (2021) 656–680.[25] B. Kitchenham, Procedures for performing systematic reviews, Keele, UK, Keele University 33 (2004) (2004)1–26.[26] ˚A. L. Hauge, J. Thomsen, T. Berker, User evaluations of energy efficient buildings: Literature review and furtherresearch, Advances in Building Energy Research 5 (1) (2011) 109–127.[27] J. Taherahmadi, Y. Noorollahi, M. Panahi, Toward comprehensive zero energy building definitions: a literaturereview and recommendations, International Journal of Sustainable Energy (2020) 1–29.[28] A. Boodi, K. Beddiar, M. Benamour, Y. Amirat, M. Benbouzid, Intelligent systems for building energy andoccupant comfort optimization: A state of the art review and recommendations, Energies 11 (10) (2018) 2604.[29] A. De Paola, M. Ortolani, G. Lo Re, G. Anastasi, S. K. Das, Intelligent management systems for energy effi-ciency in buildings: A survey, ACM Computing Surveys (CSUR) 47 (1) (2014) 1–38.[30] I. Khajenasiri, A. Estebsari, M. Verhelst, G. Gielen, A review on internet of things solutions for intelligentenergy control in buildings for smart city applications, Energy Procedia 111 (2017) 770–779.[31] H. Shareef, M. S. Ahmed, A. Mohamed, E. Al Hassan, Review on home energy management system consideringdemand responses, smart technologies, and intelligent controllers, Ieee Access 6 (2018) 24498–24509.[32] P. Resnick, H. R. Varian, Recommender systems, Communications of the ACM 40 (3) (1997) 56–58.[33] J. B. Schafer, J. Konstan, J. Riedl, Recommender systems in e-commerce, in: Proceedings of the 1st ACMconference on Electronic commerce, 1999, pp. 158–166.[34] F. J. Martin, Recsys’ 09 industrial keynote: top 10 lessons learned developing deploying and operating real-world recommender systems, in: Proceedings of the third ACM conference on Recommender systems, 2009,pp. 1–2. 25
PREPRINT - F
EBRUARY
16, 2021[35] J. Bao, Y. Zheng, D. Wilkie, M. Mokbel, Recommendations in location-based social networks: a survey, GeoIn-formatica 19 (3) (2015) 525–565.[36] A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, C. Sardianos, I. Varlamis, G. Dimitrakopoulos, Achieving do-mestic energy efficiency using micro-moments and intelligent recommendations, IEEE Access 8 (2020) 15047–15055.[37] C. Sardianos, I. Varlamis, G. Dimitrakopoulos, D. Anagnostopoulos, A. Alsalemi, F. Bensaali, A. Amira, ”i want to... change”: Micro-moment based recommendations can change users’ energy habits., in: SMART-GREENS, 2019, pp. 30–39.[38] M. Eirinaki, J. Gao, I. Varlamis, K. Tserpes, Recommender systems for large-scale social networks: A reviewof challenges and solutions, Future Generation Computer Systems 78 (2018) 413–418.[39] P. Castells, N. J. Hurley, S. Vargas, Novelty and diversity in recommender systems, in: Recommender systemshandbook, Springer, 2015, pp. 881–918.[40] T. Pinto, R. Faia, M. Navarro, G. Santos, J. Corchado Rodr´ıguez, Z. Vale, Multi-agent-based cbr recommendersystem for intelligent energy management in buildings, IEEE Systems Journal PP (2018) 1–12.[41] E. Kaur, S. Sharma, A. Verma, M. Singh, An energy management and recommender system for lighting controlin internet-of-energy enabled buildings, IFAC-PapersOnLine 52 (4) (2019) 288–293.[42] C. Cui, T. Wu, M. Hu, J. D. Weir, X. Li, Short-term building energy model recommendation system: A meta-learning approach, Applied energy 172 (2016) 251–263.[43] G. Cuffaro, F. Paganelli, G. Mylonas, A resource-based rule engine for energy savings recommendations ineducational buildings, in: 2017 Global Internet of Things Summit (GIoTS), IEEE, 2017, pp. 1–6.[44] P. Wei, S. Xia, X. Jiang, Energy saving recommendations and user location modeling in commercial build-ings, in: Proceedings of the 26th Conference on User Modeling, Adaptation and Personalization, UMAP ’18,Association for Computing Machinery, ACM, New York, NY, USA, 2018, p. 3–11.[45] P. Wei, S. Xia, R. Chen, J. Qian, C. Li, X. Jiang, A deep reinforcement learning based recommender system foroccupant-driven energy optimization in commercial buildings, IEEE Internet of Things Journal.[46] P. Kar, A. Shareef, A. Kumar, K. T. Harn, B. Kalluri, S. K. Panda, Revicee: A recommendation based approachfor personalized control, visual comfort & energy efficiency in buildings, Building and Environment 152 (2019)135 – 144.[47] D. Schweizer, M. Zehnder, H. Wache, H. Witschel, D. Zanatta, M. Rodriguez, Using consumer behavior datato reduce energy consumption in smart homes: Applying machine learning to save energy without loweringcomfort of inhabitants, in: 2015 IEEE 14th International Conference on Machine Learning and Applications(ICMLA), 2015, pp. 1123–1129.[48] T. Osadchiy, I. Poliakov, P. Olivier, M. Rowland, E. Foster, Recommender system based on pairwise associationrules, Expert Systems with Applications 115 (2019) 535 – 542.[49] J. Dahihande, A. Jaiswal, A. A. Pagar, A. Thakare, M. Eirinaki, I. Varlamis, Reducing energy waste in house-holds through real-time recommendations, in: Fourteenth ACM Conference on Recommender Systems, 2020,pp. 545–550.[50] J. Morawski, T. Stepan, S. Dick, J. Miller, A fuzzy recommender system for public library catalogs, Interna-tional Journal of Intelligent Systems 32 (10) (2017) 1062–1084.[51] J. Castro, M. J. Barranco, R. M. Rodr´ıguez, L. Mart´ınez, Group recommendations based on hesitant fuzzy sets,International Journal of Intelligent Systems 33 (10) (2018) 2058–2077.[52] Y. Zhang, K. Meng, W. Kong, Z. Y. Dong, Collaborative filtering-based electricity plan recommender system,IEEE Transactions on Industrial Informatics 15 (3) (2019) 1393–1404.[53] J. Zheng, C. S. Lai, H. Yuan, Z. Y. Dong, K. Meng, L. L. Lai, Electricity plan recommender system withelectrical instruction-based recovery, Energy 203 (2020) 117775.[54] G. Adomavicius, A. Tuzhilin, Context-aware recommender systems, in: Recommender systems handbook,Springer, 2011, pp. 217–253.[55] C. Sardianos, I. Varlamis, C. Chronis, G. Dimitrakopoulos, Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Amodel for predicting room occupancy based on motion sensor data, in: 2020 IEEE International Conference onInformatics, IoT, and Enabling Technologies (ICIoT), IEEE, 2020, pp. 394–399.[56] S. Raza, C. Ding, Progress in context-aware recommender systems — an overview, Computer Science Review31 (2019) 84 – 97. 26
PREPRINT - F
EBRUARY
16, 2021[57] A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, C. Sardianos, C. Chronis, I. Varlamis, G. Dimitrakopoulos, Amicro-moment system for domestic energy efficiency analysis, IEEE Systems Journal.[58] H. Shigeyoshi, K. Tamano, R. Saga, H. Tsuji, S. Inoue, T. Ueno, Social experiment on advisory recommendersystem for energy-saving, in: S. Yamamoto (Ed.), Human Interface and the Management of Information. Infor-mation and Interaction Design, Springer, Springer Berlin Heidelberg, Berlin, Heidelberg, 2013, pp. 545–554.[59] M. Radha, M. C. Willemsen, M. Boerhof, W. A. IJsselsteijn, Lifestyle recommendations for hypertensionthrough rasch-based feasibility modeling, in: Proceedings of the 2016 Conference on User Modeling Adap-tation and Personalization, 2016, pp. 239–247.[60] A. Starke, M. Willemsen, C. Snijders, Saving energy in 1-d: Tailoring energy-saving advice using a Rasch-based energy recommender system, in: Proceedings of 2nd International Workshop on Decision Making andRecommender Systems, CEUR Workshop Proceedings, 2015, pp. 1–8.[61] A. Starke, M. Willemsen, C. Snijders, Effective user interface designs to increase energy-efficient behavior in arasch-based energy recommender system, in: Proceedings of the Eleventh ACM Conference on RecommenderSystems, RecSys ’17, ACM, Association for Computing Machinery, New York, NY, USA, 2017, p. 65–73.[62] R. Chulyadyo, P. Leray, A personalized recommender system from probabilistic relational model and users’preferences, Procedia Computer Science 35 (2014) 1063–1072.[63] R. Kumar, P. Raghavan, S. Rajagopalan, A. Tomkins, Recommendation systems: A probabilistic analysis,Journal of Computer and System Sciences 63 (1) (2001) 42–61.[64] L. Kung, H.-F. Wang, A recommender system for the optimal combination of energy resources with cost-benefitanalysis, in: 2015 International Conference on Industrial Engineering and Operations Management (IEOM),IEEE, 2015, pp. 1–10.[65] K. Oku, F. Hattori, Fusion-based recommender system for improving serendipity, CEUR Workshop Proceedings816 (2011) 19–26.[66] X. Xin, Effective fusion-based approaches for recommender systems, Chinese University of Hong Kong, 2011.[67] K. Zhang, H. Li, Fusion-based recommender system, in: 2010 13th International Conference on InformationFusion, IEEE, 2010, pp. 1–7.[68] Y. Himeur, A. Alsalemi, A. Al-Kababji, F. Bensaali, A. Amira, Data fusion strategies for energy efficiency inbuildings: Overview, challenges and novel orientations, Information Fusion 64 (2020) 99–120.[69] A. Wroblewska, J. Dabrowski, M. Pastuszak, A. Michalowski, M. Daniluk, B. Rychalska, M. Wieczorek,S. Sysko-Romanczuk, Multi-modal embedding fusion-based recommender (2020). arXiv:2005.06331 .[70] Z. Ji, C. Yang, H. Wang, J. E. Armend´ariz-i˜nigo, M. Arce-Urriza, Brs c s: a hybrid recommendation model fus-ing multi-source heterogeneous data, EURASIP Journal on Wireless Communications and Networking 2020 (1)(2020) 1–17.[71] Q. Shambour, J. Lu, A trust-semantic fusion-based recommendation approach for e-business applications, De-cision Support Systems 54 (1) (2012) 768–780.[72] H. Wang, Y. Song, P. Mi, J. Duan, The collaborative filtering method based on social information fusion,Mathematical Problems in Engineering 2019.[73] T. Pradhan, S. Pal, A multi-level fusion based decision support system for academic collaborator recommenda-tion, Knowledge-Based Systems (2020) 105784.[74] G. Aiello, I. Giovino, M. Vallone, P. Catania, A. Argento, A decision support system based on multisensor datafusion for sustainable greenhouse management, Journal of Cleaner Production 172 (2018) 4057–4065.[75] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, A novel approach for detecting anomalous energy consumptionbased on micro-moments and deep neural networks, Cognitive Computation 12 (6) (2020) 1381–1401.[76] S. Zhang, L. Yao, A. Sun, Y. Tay, Deep learning based recommender system: A survey and new perspectives,ACM Computing Surveys (CSUR) 52 (1) (2019) 1–38.[77] J. Bobadilla, S. Alonso, A. Hernando, Deep learning architecture for collaborative filtering recommender sys-tems, Applied Sciences 10 (7).[78] K. R, P. Kumar, B. Bhasker, Dnnrec: A novel deep learning based hybrid recommender system, Expert Systemswith Applications 144 (2020) 113054.[79] R. Lara-Cabrera, ´A. Gonz´alez-Prieto, F. Ortega, Deep matrix factorization approach for collaborative filteringrecommender systems, Applied Sciences 10 (14) (2020) 4926.27
PREPRINT - F
EBRUARY
16, 2021[80] A. S. Shah, H. Nasir, M. Fayaz, A. Lajis, A. Shah, A review on energy consumption optimization techniques iniot based smart building environments, Information 10 (3) (2019) 108.[81] Q. Tushar, M. Bhuiyan, M. Sandanayake, G. Zhang, Optimizing the energy consumption in a residential build-ing at different climate zones: Towards sustainable decision making, Journal of cleaner production 233 (2019)634–649.[82] I. Ceballos-Fuentealba, E. ´Alvarez-Miranda, C. Torres-Fuchslocher, M. L. del Campo-Hitschfeld, J. D´ıaz-Guerrero, A simulation and optimisation methodology for choosing energy efficiency measures in non-residential buildings, Applied Energy 256 (2019) 113953.[83] P. Rocha, A. Siddiqui, M. Stadler, Improving energy efficiency via smart building energy management systems:A comparison with policy measures, Energy and Buildings 88 (2015) 203–213.[84] A. Anvari-Moghaddam, A. Rahimi-Kian, M. S. Mirian, J. M. Guerrero, A multi-agent based energy manage-ment solution for integrated buildings and microgrid system, Applied energy 203 (2017) 41–56.[85] S. Lu, W. Gu, K. Meng, Z. Y. Dong, Economic dispatch of integrated energy systems with robust thermalcomfort management, IEEE Transactions on Sustainable Energy.[86] M. Di Piazza, G. La Tona, M. Luna, A. Di Piazza, A two-stage energy management system for smart buildingsreducing the impact of demand uncertainty, Energy and Buildings 139 (2017) 1–9.[87] S. Salakij, N. Yu, S. Paolucci, P. Antsaklis, Model-based predictive control for building energy management. i:Energy modeling and optimal control, Energy and Buildings 133 (2016) 345–358.[88] N. Yu, S. Salakij, R. Chavez, S. Paolucci, M. Sen, P. Antsaklis, Model-based predictive control for buildingenergy management: Part ii–experimental validations, Energy and Buildings 146 (2017) 19–26.[89] S. Paul, N. P. Padhy, Real-time bilevel energy management of smart residential apartment building, IEEE Trans-actions on Industrial Informatics 16 (6) (2019) 3708–3720.[90] S. Lu, W. Gu, K. Meng, S. Yao, B. Liu, Z. Y. Dong, Thermal inertial aggregation model for integrated energysystems, IEEE Transactions on Power Systems 35 (3) (2019) 2374–2387.[91] A. A. Abbasi, A. Abbasi, S. Shamshirband, A. T. Chronopoulos, V. Persico, A. Pescap`e, Software-defined cloudcomputing: A systematic review on latest trends and developments, IEEE Access 7 (2019) 93294–93314.[92] A. Elsalemi, A. Al-kababji, Y. Himeur, F. Bensaali, A. Amira, Cloud energy micro-moment data classification:A platform study, in: IEEE SmartWorld, Ubiquitous Intelligence Computing, Advanced Trusted Computing,Scalable Computing Communications, Cloud Big Data Computing, Internet of People and Smart City Innova-tion (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), 2020, pp. 1–6.[93] J. Zhou, Z. Cao, X. Dong, A. V. Vasilakos, Security and privacy for cloud-based iot: Challenges, IEEE Com-munications Magazine 55 (1) (2017) 26–33.[94] J. Ren, D. Zhang, S. He, Y. Zhang, T. Li, A survey on end-edge-cloud orchestrated network computingparadigms: Transparent computing, mobile edge computing, fog computing, and cloudlet, ACM ComputingSurveys (CSUR) 52 (6) (2019) 1–36.[95] A. Elsalemi, Y. Himeur, F. Bensaali, A. Amira, Appliance-level monitoring with micro-moment smart plugs,in: The Fifth International Conference on Smart City Applications (SCA), 2020, pp. 1–5.[96] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Effective non-intrusive load monitoring of buildings based on anovel multi-descriptor fusion with dimensionality reduction, Applied Energy 279 (2020) 115872.[97] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, Appliance identification using a histogram post-processing of 2dlocal binary patterns for smart grid applications, in: Proc. 25th International Conference on Pattern Recognition(ICPR), 2020, pp. 1–8.[98] M. Devarajan, V. Subramaniyaswamy, V. Vijayakumar, L. Ravi, Fog-assisted personalized healthcare-supportsystem for remote patients with diabetes, Journal of Ambient Intelligence and Humanized Computing 10 (10)(2019) 3747–3760.[99] E. Hern´andez-Nieves, G. Hern´andez, A.-B. Gil-Gonz´alez, S. Rodr´ıguez-Gonz´alez, J. M. Corchado, Fog com-puting architecture for personalized recommendation of banking products, Expert Systems with Applications140 (2020) 112900.[100] X. Wang, B. Gu, Y. Ren, W. Ye, S. Yu, Y. Xiang, L. Gao, A fog-based recommender system, IEEE Internet ofThings Journal 7 (2) (2019) 1048–1060.[101] D. Linthicum, Edge computing in hybrid cloud: 3 approaches (2019). https://techbeacon.com/enterprise-it/edge-computing-hybrid-cloud-3-approaches PREPRINT - F
EBRUARY
16, 2021[102] Y. Huang, F. Wang, F. Wang, J. Liu, Deepar: A hybrid device-edge-cloud execution framework for mobile deeplearning applications, in: IEEE INFOCOM 2019-IEEE Conference on Computer Communications Workshops(INFOCOM WKSHPS), IEEE, 2019, pp. 892–897.[103] A. Gunawardana, G. Shani, Evaluating Recommender Systems, in: Recommender systems handbook, Springer,2015, pp. 265–308.[104] W. Wu, L. He, J. Yang, Evaluating recommender systems, in: Seventh International Conference on DigitalInformation Management (ICDIM 2012), IEEE, 2012, pp. 56–61.[105] T. Chai, R. R. Draxler, Root mean square error (rmse) or mean absolute error (mae)?–arguments against avoid-ing rmse in the literature, Geoscientific model development 7 (3) (2014) 1247–1250.[106] J. L. Herlocker, J. A. Konstan, L. G. Terveen, J. T. Riedl, Evaluating collaborative filtering recommender sys-tems, ACM Transactions on Information Systems (TOIS) 22 (1) (2004) 5–53.[107] J. Herlocker, J. A. Konstan, J. Riedl, An empirical analysis of design choices in neighborhood-based collabora-tive filtering algorithms, Information retrieval 5 (4) (2002) 287–310.[108] S. M. McNee, J. Riedl, J. A. Konstan, Making Recommendations Better: An Analytic Model for Human-Recommender Interaction, in: CHI’06 extended abstracts on Human factors in computing systems, 2006, pp.1103–1108.[109] S. Karjalainen, Consumer preferences for feedback on household electricity consumption, Energy and buildings43 (2-3) (2011) 458–467.[110] B. Karlin, J. F. Zinger, R. Ford, The effects of feedback on energy conservation: A meta-analysis., Psychologicalbulletin 141 (6) (2015) 1205.[111] A. Garbi, A. Malamou, N. Michas, Z. Pontikas, N. Doulamis, E. Protopapadakis, T. N. Mikkelsen, K. Kanel-lakis, J.-L. Baradat, Beneffice: Behaviour change, consumption monitoring and analytics with complementarycurrency rewards, Proceedings 20 (1) (2019) 12.[112] P. Petkov, F. K¨obler, M. Foth, R. Medland, H. Krcmar, Engaging energy saving through motivation-specificsocial comparison, in: CHI’11 Extended Abstracts on Human Factors in Computing Systems, Association forComputing Machinery, 2011, pp. 1945–1950.[113] R. K. Jain, R. Gulbinas, J. E. Taylor, P. J. Culligan, Can social influence drive energy savings? detectingthe impact of social influence on the energy consumption behavior of networked users exposed to normativeeco-feedback, Energy and Buildings 66 (2013) 119–127.[114] F. Du, J. Zhang, H. Li, J. Yan, S. Galloway, K. L. Lo, Modelling the impact of social network on energy savings,Applied Energy 178 (2016) 56–65.[115] D. Wemyss, F. Cellina, E. Lobsiger-K¨agi, V. De Luca, R. Castri, Does it last? long-term impacts of an app-based behavior change intervention on household electricity savings in switzerland, Energy Research & SocialScience 47 (2019) 16–27.[116] J. Morley, K. Widdicks, M. Hazas, Digitalisation, energy and data demand: The impact of internet traffic onoverall and peak electricity consumption, Energy Research & Social Science 38 (2018) 128–137.[117] O. Garcia, J. Prieto, R. Alonso, J. Corchado Rodr´ıguez, A framework to improve energy efficient behaviour athome through activity and context monitoring, Sensors 17 (2017) 1749.[118] F. Luo, G. Ranzi, W. Kong, Z. Dong, S. Wang, J. Zhao, Non-intrusive energy saving appliance recommendersystem for smart grid residential users, IET Generation, Transmission and Distribution 11.[119] C. Sardianos, I. Varlamis, G. Dimitrakopoulos, D. Anagnostopoulos, A. Alsalemi, F. Bensaali, Y. Himeur,A. Amira, Rehab-c: Recommendations for energy habits change, Future Generation Computer Systems 112(2020) 394 – 407.[120] D. Jim´enez-Bravo, J. P´erez-Marcos, D. Hern´andez de la Iglesia, G. Villarrubia, J. Paz, Multi-agent recommen-dation system for electrical energy optimization and cost saving in smart homes, Energies 12 (2019) 1317.[121] A. A. A. Ari, O. K. Ngangmo, C. Titouna, O. Thiare, A. Mohamadou, A. M. Gueroui, et al., Enabling privacyand security in cloud of things: Architecture, applications, security & privacy challenges, Applied Computingand Informatics.[122] J. L. Schaefer, J. C. M. Siluk, P. S. d. Carvalho, J. Renes Pinheiro, P. S. Schneider, Management challenges andopportunities for energy cloud development and diffusion, Energies 13 (16) (2020) 4048.[123] T. Mahjabin, Y. Xiao, G. Sun, W. Jiang, A survey of distributed denial-of-service attack, prevention, and miti-gation techniques, International Journal of Distributed Sensor Networks 13 (12) (2017) 1550147717741463.29
PREPRINT - F
EBRUARY
16, 2021[124] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, I. Varlamis, G. Bravos, C. Sardianos, Dimitrakopoulos, Techno-economic analysis of building energy efficiency systems based on behavioral change: A case study of a novelmicro-moments based solution, Applied Energy (2021) 1–25.[125] S. Natarajan, S. Vairavasundaram, S. Natarajan, A. H. Gandomi, Resolving data sparsity and cold start problemin collaborative filtering recommender system using linked open data, Expert Systems with Applications 149(2020) 113248.[126] A. F. Jain, S. K. Vishwakarma, P. Jain, An efficient collaborative recommender system for removing sparsityproblem, in: ICT Analysis and Applications, Springer, 2020, pp. 131–141.[127] F. Zhang, S. Qi, Q. Liu, M. Mao, A. Zeng, Alleviating the data sparsity problem of recommender systems byclustering nodes in bipartite networks, Expert Systems with Applications (2020) 113346.[128] L. H. Son, Dealing with the new user cold-start problem in recommender systems: A comparative review,Information Systems 58 (2016) 87–104.[129] B. Lika, K. Kolomvatsos, S. Hadjiefthymiades, Facing the cold start problem in recommender systems, ExpertSystems with Applications 41 (4) (2014) 2065–2073.[130] J.-H. Liu, T. Zhou, Z.-K. Zhang, Z. Yang, C. Liu, W.-M. Li, Promoting cold-start items in recommender sys-tems, PloS one 9 (12) (2014) e113457.[131] K. Verbert, H. Drachsler, N. Manouselis, M. Wolpers, R. Vuorikari, E. Duval, Dataset-driven research forimproving recommender systems for learning, in: Proceedings of the 1st International Conference on LearningAnalytics and Knowledge, 2011, pp. 44–53.[132] H. Drachsler, T. Bogers, R. Vuorikari, K. Verbert, E. Duval, N. Manouselis, G. Beham, S. Lindstaedt, H. Stern,M. Friedrich, et al., Issues and considerations regarding sharable data sets for recommender systems in technol-ogy enhanced learning, Procedia Computer Science 1 (2) (2010) 2849–2858.[133] E. C¸ ano, M. Morisio, Characterization of public datasets for recommender systems, in: 2015 IEEE 1st Inter-national Forum on Research and Technologies for Society and Industry Leveraging a better tomorrow (RTSI),IEEE, 2015, pp. 249–257.[134] F. Isinkaye, Y. Folajimi, B. A. Ojokoh, Recommendation systems: Principles, methods and evaluation, Egyptianinformatics journal 16 (3) (2015) 261–273.[135] M. F. Dacrema, S. Boglio, P. Cremonesi, D. Jannach, A troubling analysis of reproducibility and progress inrecommender systems research, arXiv preprint arXiv:1911.07698.[136] M. D. Ekstrand, M. Ludwig, J. A. Konstan, J. T. Riedl, Rethinking the recommender research ecosystem:reproducibility, openness, and lenskit, in: Proceedings of the fifth ACM conference on Recommender systems,2011, pp. 133–140.[137] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Building power consumption datasets: Survey, taxonomy andfuture directions, Energy and Buildings 227 (2020) 110404.[138] J. Beel, C. Breitinger, S. Langer, A. Lommatzsch, B. Gipp, Towards reproducibility in recommender-systemsresearch, User modeling and user-adapted interaction 26 (1) (2016) 69–101.[139] E. Ie, C. wei Hsu, M. Mladenov, V. Jain, S. Narvekar, J. Wang, R. Wu, C. Boutilier, Recsim: A configurablesimulation platform for recommender systems (2019). arXiv:1909.04847 .[140] A. Trianni, E. Cagno, F. Marchesani, G. Spallina, Drivers for industrial energy efficiency: an innovative ap-proach, in: ICAE–International Conference on Applied Energy, 2013, pp. 1–9.[141] O. Fitch-Roy, J. Fairbrass, Negotiating the EU’s 2030 climate and energy framework: Agendas, ideas andEuropean interest groups, Springer, 2018.[142] F. Kaabi, Conservation plan for tarsheed, Qatar General Electricity and Water Corporation, Conservation &Energy Efficiency Department, Doha, Qatar.[143] Qatar General Electricity & water Corporation, Available online: , note = Accessed: 2020-11-05,.[144] D. Gunning, Explainable artificial intelligence (xai), Defense Advanced Research Projects Agency (DARPA),nd Web 2 (2).[145] A. B. Arrieta, N. D´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc´ıa, S. Gil-L´opez,D. Molina, R. Benjamins, et al., Explainable artificial intelligence (xai): Concepts, taxonomies, opportunitiesand challenges toward responsible ai, Information Fusion 58 (2020) 82–115.30
PREPRINT - F
EBRUARY
16, 2021[146] Y. Wu, Z. Zhang, G. Kou, H. Zhang, X. Chao, C.-C. Li, Y. Dong, F. Herrera, Distributed linguistic repre-sentations in decision making: Taxonomy, key elements and applications, and challenges in data science andexplainable artificial intelligence, Information Fusion 65 (2021) 165 – 178.[147] Y. Zhang, X. Chen, et al., Explainable recommendation: A survey and new perspectives, Foundations andTrends® in Information Retrieval 14 (1) (2020) 1–101.[148] O. Sagi, L. Rokach, Explainable decision forest: Transforming a decision forest into an interpretable tree,Information Fusion 61 (2020) 124 – 138.[149] R. R. Fern´andez, I. M. de Diego, V. Ace˜na, A. Fern´andez-Isabel, J. M. Moguerza, Random forest explainabilityusing counterfactual sets, Information Fusion 63 (2020) 196–207.[150] Y. Zhang, X. Chen, Explainable recommendation: A survey and new perspectives, arXiv preprintarXiv:1804.11192.[151] J. Gao, X. Wang, Y. Wang, X. Xie, Explainable recommendation through attentive multi-view learning, in:Proceedings of the 33rd Conference on Artificial Intelligence, Vol. 33, AAAI, 2019, pp. 3622–3629.[152] K. Balog, F. Radlinski, S. Arakelyan, Transparent, scrutable and explainable user models for personalizedrecommendation, in: Proceedings of the 42nd International ACM SIGIR Conference on Research and Devel-opment in Information Retrieval, ACM, 2019, pp. 265–274.[153] R. Catherine, K. Mazaitis, M. Eskenazi, W. Cohen, Explainable entity-based recommendations with knowledgegraphs, arXiv preprint arXiv:1707.05254.[154] Y. Zhang, G. Lai, M. Zhang, Y. Zhang, Y. Liu, S. Ma, Explicit factor models for explainable recommendationbased on phrase-level sentiment analysis, in: Proceedings of the 37th International ACM SIGIR conference onResearch & Development in Information Retrieval, ACM, 2014, pp. 83–92.[155] X. He, T. Chen, M.-Y. Kan, X. Chen, Trirank: Review-aware explainable recommendation by modeling aspects,in: Proceedings of the 24th ACM International on Conference on Information and Knowledge Management,ACM, 2015, pp. 1661–1670.[156] A. I. Grimaldo, J. Novak, User-centered visual analytics approach for interactive and explainable energy demandanalysis in prosumer scenarios, in: International Conference on Computer Vision Systems, Springer, 2019, pp.700–710.[157] C. Sardianos, I. Varlamis, C. Chronis, G. Dimitrakopoulos, Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Dataanalytics, automations, and micro-moment based recommendations for energy efficiency, in: 2020 IEEE SixthInternational Conference on Big Data Computing Service and Applications (BigDataService), IEEE, 2020, pp.96–103.[158] F. Shi, J.-L. Marini, E. Audry, Towards a psycho-cognitive recommender system, in: Proceedings of the Inter-national Workshop on Emotion Representations and Modelling for Companion Technologies, 2015, pp. 25–31.[159] L. Zhao, J. Huang, N. Zhong, A context-aware recommender system with a cognition inspired model, in:International Conference on Rough Sets and Knowledge Technology, Springer, 2014, pp. 613–622.[160] P. Shafto, O. Nasraoui, Human-recommender systems: From benchmark data to benchmark cognitive models,in: Proceedings of the 10th ACM Conference on Recommender Systems, 2016, pp. 127–130.[161] S. Kopeinik, D. Kowald, I. Hasani-Mavriqi, E. Lex, Improving collaborative filtering using a cognitive modelof human category learning, The Journal of Web Science 2.[162] K. G. HamlAbadi, A. M. Saghiri, M. Vahdati, M. D. TakhtFooladi, M. R. Meybodi, A framework for cognitiverecommender systems in the internet of things (iot), in: 2017 IEEE 4th International Conference on Knowledge-Based Engineering and Innovation (KBEI), IEEE, 2017, pp. 0971–0976.[163] S. Kopeinik, Applying cognitive learner models for recommender systems in sparse data learning environments,SIGIR Forum 51 (3) (2018) 165.[164] J. Aguilar, P. Valdiviezo-D´ıaz, G. Riofrio, A general framework for intelligent recommender systems, Appliedcomputing and informatics 13 (2) (2017) 147–160.[165] A. Beheshti, S. Yakhchi, S. Mousaeirad, S. M. Ghafari, S. R. Goluguri, M. A. Edrisi, Towards cognitive recom-mender systems, Algorithms 13 (8) (2020) 176.[166] E. Bothos, D. Apostolou, G. Mentzas, A recommender for persuasive messages in route planning applications,in: 2016 7th International Conference on Information, Intelligence, Systems & Applications (IISA), IEEE,2016, pp. 1–5. 31
PREPRINT - F
EBRUARY
16, 2021[167] R. S´anchez-Corcuera, D. Casado-Mansilla, C. E. Borges, D. L´opez-de Ipi˜na, Persuasion-based recommendersystem ensambling matrix factorisation and active learning models, Personal and Ubiquitous Computing (2020)1–11.[168] S. Puglisi, J. Parra-Arnau, J. Forn´e, D. Rebollo-Monedero, On content-based recommendation and user privacyin social-tagging systems, Computer Standards & Interfaces 41 (2015) 17–27.[169] Q. Tang, H. Wang, Privacy-preserving hybrid recommender system, in: Proceedings of the Fifth ACM Interna-tional Workshop on Security in Cloud Computing, 2017, pp. 59–66.[170] H. Shin, S. Kim, J. Shin, X. Xiao, Privacy enhanced matrix factorization for recommendation with local differ-ential privacy, IEEE Transactions on Knowledge and Data Engineering 30 (9) (2018) 1770–1782.[171] D. Li, Q. Lv, L. Shang, N. Gu, Efficient privacy-preserving content recommendation for online social commu-nities, Neurocomputing 219 (2017) 440–454.[172] C.-M. Tseng, C.-K. Chau, On the privacy of crowd-sourced data collection for distance-to-empty prediction andeco-routing, in: Proceedings of the Workshop on Electric Vehicle Systems, Data, and Applications, 2016, pp.1–6.[173] X. Wang, J. Zhang, Y. Wang, Trust-aware privacy-preserving recommender system, in: Proceedings of the 9thEAI International Conference on Mobile Multimedia Communications, 2016, pp. 107–115.[174] S. McLaughlin, P. McDaniel, W. Aiello, Protecting consumer privacy from electric load monitoring, in: Pro-ceedings of the 18th ACM conference on Computer and communications security, 2011, pp. 87–98.[175] Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira, Smart non-intrusive appliance identification using a novel localpower histogramming descriptor with an improved k-nearest neighbors classifier, Sustainable Cities and Society67 (2021) 102764.[176] F. Englert, M. Rettberg-P¨aplow, S. K¨ossler, A. Alhamoud, T. A. B. Nguyen, D. B¨ohnstedt, R. Steinmetz, En-hancing user privacy by data driven selection mechanisms for finding transmission-relevant data samples inenergy recommender systems, in: 2015 International Conference and Workshops on Networked Systems (Net-Sys), IEEE, 2015, pp. 1–6.[177] S. Badsha, X. Yi, I. Khalil, E. Bertino, Privacy preserving user-based recommender system, in: 2017 IEEE 37thinternational conference on Distributed Computing Systems (ICDCS), IEEE, 2017, pp. 1074–1083.[178] J.-Y. Jiang, C.-T. Li, S.-D. Lin, Towards a more reliable privacy-preserving recommender system, InformationSciences 482 (2019) 248–265.[179] S. Badsha, X. Yi, I. Khalil, A practical privacy-preserving recommender system, Data Science and Engineering1 (3) (2016) 161–177.[180] K. Xu, W. Zhang, Z. Yan, A privacy-preserving mobile application recommender system based on trust evalua-tion, Journal of computational science 26 (2018) 87–107.[181] N. Begum, M. Z. A. Bhuiyan, Towards privacy-preserving recommender system with blockchains, in: Depend-ability in Sensor, Cloud, and Big Data Systems and Applications: 5th International Conference, DependSys2019, Guangzhou, China, November 12–15, 2019, Proceedings, Vol. 1123, Springer Nature, 2019, p. 106.[182] Y. Pu, T. Xiang, C. Hu, A. Alrawais, H. Yan, An efficient blockchain-based privacy preserving scheme forvehicular social networks, Information Sciences 540 (2020) 308–324.[183] R. Bosri, M. S. Rahman, M. Z. A. Bhuiyan, A. Al Omar, Integrating blockchain with artificial intelligence forprivacy-preserving in recommender systems, IEEE Transactions on Network Science and Engineering.[184] F. Casino, C. Patsakis, An efficient blockchain-based privacy-preserving collaborative filtering architecture,IEEE Transactions on Engineering Management.[185] Z. Gantner, S. Rendle, L. Schmidt-Thieme, Factorization models for context-/time-aware movie recommenda-tions, in: Proceedings of the Workshop on Context-Aware Movie Recommendation, 2010, pp. 14–19.[186] Q. Yuan, G. Cong, Z. Ma, A. Sun, N. M. Thalmann, Time-aware point-of-interest recommendation, in: Proceed-ings of the 36th international ACM SIGIR conference on Research and development in information retrieval,2013, pp. 363–372.[187] J.-D. Zhang, C.-Y. Chow, Ticrec: A probabilistic framework to utilize temporal influence correlations for time-aware location recommendations, IEEE Transactions on Services Computing 9 (4) (2015) 633–646.[188] K. Stefanidis, E. Ntoutsi, M. Petropoulos, K. Nørv˚ag, H.-P. Kriegel, A framework for modeling, computingand presenting time-aware recommendations, in: Transactions on Large-Scale Data-and Knowledge-CenteredSystems X, Springer, 2013, pp. 146–172. 32
PREPRINT - F
EBRUARY
16, 2021[189] P. G. Campos, F. D´ıez, I. Cantador, Time-aware recommender systems: a comprehensive survey and analysisof existing evaluation protocols, User Modeling and User-Adapted Interaction 24 (1-2) (2014) 67–119.[190] S. Linda, S. Minz, K. Bharadwaj, Effective context-aware recommendations based on context weighting usinggenetic algorithm and alleviating data sparsity, Applied Artificial Intelligence 34 (10) (2020) 730–753.[191] K. Wang, Y. Jin, H. Wang, H. Peng, X. Wang, Personalized time-aware tag recommendation, in: The Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18), 2018, pp. 459–466.[192] Y. Qian, Y. Zhang, X. Ma, H. Yu, L. Peng, Ears: Emotion-aware recommender system based on hybrid infor-mation fusion, Information Fusion 46 (2019) 141–146.[193] M. Nilashi, S. Asadi, R. A. Abumalloh, S. Samad, O. Ibrahim, Intelligent recommender systems in the covid-19outbreak: The case of wearable healthcare devices, Journal of Soft Computing and Decision Support Systems7 (4) (2020) 8–12.[194] B. Lamche, Y. R¨odl, C. Hauptmann, W. W¨orndl, Context-aware recommendations for mobile shopping., in:LocalRec@ RecSys, 2015, pp. 21–27.[195] M. Unger, A. Tuzhilin, A. Livne, Context-aware recommendations based on deep learning frameworks, ACMTransactions on Management Information Systems (TMIS) 11 (2) (2020) 1–15.[196] P. S´anchez, A. Bellog´ın, Time-aware novelty metrics for recommender systems, in: European Conference onInformation Retrieval, Springer, 2018, pp. 357–370.[197] M. Kunaver, T. Poˇzrl, Diversity in recommender systems–a survey, Knowledge-Based Systems 123 (2017)154–162.[198] Z. Batmaz, A. Yurekli, A. Bilge, C. Kaleli, A review on deep learning for recommender systems: challengesand remedies, Artificial Intelligence Review 52 (1) (2019) 1–37.[199] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, J. Leskovec, Graph convolutional neural networksfor web-scale recommender systems, in: Proceedings of the 24th ACM SIGKDD International Conference onKnowledge Discovery & Data Mining, 2018, pp. 974–983.[200] C.-L. Liu, X.-W. Wu, Large-scale recommender system with compact latent factor model, Expert Systems withApplications 64 (2016) 467–475.[201] A. Barraza-Urbina, D. Glowacka, Introduction to bandits in recommender systems, in: Fourteenth ACM Con-ference on Recommender Systems, 2020, pp. 748–750.[202] Q. Zhou, X. Zhang, J. Xu, B. Liang, Large-scale bandit approaches for recommender systems, in: InternationalConference on Neural Information Processing, Springer, 2017, pp. 811–821.[203] R. Ca˜namares, M. Redondo, P. Castells, Multi-armed recommender system bandit ensembles, in: Proceedingsof the 13th ACM Conference on Recommender Systems, 2019, pp. 432–436.[204] J. Sanz-Cruzado, P. Castells, E. L´opez, A simple multi-armed nearest-neighbor bandit for interactive recom-mendation, in: Proceedings of the 13th ACM Conference on Recommender Systems, 2019, pp. 358–362.[205] C. Sardianos, I. Varlamis, M. Eirinaki, Scaling collaborative filtering to large–scale bipartite rating graphsusing lenskit and spark, in: 2017 IEEE Third International Conference on Big Data Computing Service andApplications (BigDataService), IEEE, 2017, pp. 70–79.[206] G. Bathla, H. Aggarwal, R. Rani, Scalable recommendation using large scale graph partitioning with pregel andgiraph, International Journal of Cognitive Informatics and Natural Intelligence (IJCINI) 14 (4) (2020) 42–61.[207] H. Li, K. Li, J. An, W. Zheng, K. Li, An efficient manifold regularized sparse non-negative matrix factorizationmodel for large-scale recommender systems on gpus, Information Sciences 496 (2019) 464–484.[208] H. Li, K. Li, J. An, K. Li, Msgd: A novel matrix factorization approach for large-scale collaborative filteringrecommender systems on gpus, IEEE Transactions on Parallel and Distributed Systems 29 (7) (2017) 1530–1544.[209] C. Sardianos, N. Tsirakis, I. Varlamis, A survey on the scalability of recommender systems for social networks,in: Social Networks Science: Design, Implementation, Security, and Challenges, Springer, 2018, pp. 89–110.[210] Y. Zhang, W. Kong, Z. Y. Dong, K. Meng, J. Qiu, Big data-driven electricity plan recommender system, in:2018 IEEE Power & Energy Society General Meeting (PESGM), IEEE, 2018, pp. 1–5.[211] P. Kar, A. Shareef, A. Kumar, K. T. Harn, B. Kalluri, S. K. Panda, Revicee: A recommendation based approachfor personalized control, visual comfort & energy efficiency in buildings, Building and Environment 152 (2019)135–144. 33
PREPRINT - F
EBRUARY
16, 2021[212] C. Sardianos, C. Chronis, I. Varlamis, G. Dimitrakopoulos, Y. Himeur, A. Alsalemi, F. Bensaali, A. Amira,Smart fusion of sensor data and human feedback for personalised energy-saving recommendations, InternationalJournal of Intelligent Systems (2021) 1–20.[213] A. Fensel, S. Tomic, V. Kumar, M. Stefanovic, S. V. Aleshin, D. O. Novikov, SESAME-S: Semantic SmartHome System for Energy Efficiency, Informatik-Spektrum 36 (1) (2013) 46–57.[214] M. Rodr´ıguez Fern´andez, A. Cort´es Garc´ıa, I. Gonz´alez Alonso, E. Zalama Casanova, Using the Big Datagenerated by the Smart Home to improve energy efficiency management, Energy Effic. 9 (1) (2016) 249–260.[215] P. Fraternali, S. Herrera, J. Novak, M. Melenhorst, D. Tzovaras, S. Krinidis, A. E. Rizzoli, C. Rottondi, F. Cel-lina, EnCOMPASS - An Integrative Approach to Behavioural Change for Energy Saving, in: GIoTS 2017 -Glob. Internet Things Summit, Proc., 2017, pp. 1–6.[216] P. Fraternali, F. Cellina, S. Herrera, S. Krinidis, C. Pasini, A. E. Rizzoli, C. Rottondi, D. Tzovaras, A Socio-Technical System Based on Gamification Towards Energy Savings, in: 2018 IEEE Int. Conf. Pervasive Comput.Commun. Work. (PerCom Work., IEEE, 2018, pp. 59–64.[217] S. Albertarelli, P. Fraternali, J. Novak, A. E. Rizzoli, C. Rottondi, DROP and FUNERGY: Two Gamified Learn-ing Projects for Water and Energy Conservation, in: Proc. 11th Eur. Conf. Games Based Learn. ECGBL 2017,2017, pp. 935–938.[218] E. Fotopoulou, A. Zafeiropoulos, F. Terroso, A. Gonzalez, A. Skarmeta, U. Simsek, A. Fensel, Data Aggre-gation, Fusion and Recommendations for Strengthening Citizens Energy-aware Behavioural Profiles, in: 2017Glob. Internet Things Summit, 2017, pp. 1–6.[219] M. Zorrilla, ´A. Ibrain, Bernard, an energy intelligent system for raising residential users awareness, Comput.Ind. Eng. 135 (June) (2019) 492–499.[220] I. Machorro-Cano, G. Alor-Hern´andez, M. A. Paredes-Valverde, L. Rodr´ıguez-Mazahua, J. L. S´anchez-Cervantes, J. O. Olmedo-Aguirre, HEMS-IoT: A Big Data and Machine Learning-Based Smart Home Systemfor Energy Saving, Energies 13 (5).[221] P. Sitoula, D. Rahayu, P. D. Haghighi, S. Goodwin, C. Ling, Context-Aware Smart Energy Recommender(CASER), in: Proc. 17th Int. Conf. Adv. Mob. Comput. Multimed., 2019, pp. 13–19.[222] A. Alsalemi, F. Bensaali, A. Amira, N. Fetais, C. Sardianos, I. Varlamis, Smart Energy Usage and Visualizationbased on Micro-Moments, in: Adv. Intell. Syst. Comput., Vol. 1038, 2020, pp. 557–566.[223] A. Al-Kababji, A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, R. Fernandez, N. Fetais, Energy data visual-izations on smartphones for triggering behavioral change: Novel vs. conventional, in: 2020 2nd Global Power,Energy and Communication Conference (GPECOM), IEEE, 2020, pp. 312–317.[224] A. Al-Kababji, A. Alsalemi, Y. Himeur, F. Bensaali, A. Amira, R. Fernandez, N. Fetais, Interactive visualanalytics for residential energy big data, Information Visualization (2021) 1–20.[225] BBC, Coronavirus: Domestic electricity use up during day as nation works from home - BBC News (2020). [226] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, An intelligent non-intrusive load monitoring scheme based on2D phase encoding of power signals, International Journal of Intelligent Systems 36 (1) (2021) 72–93.[227] A. S. Siddiqui, A. M. Sibal, Energy Disaggregation in Smart Home Appliances: A Deep Learning Approach,Energy (2020) 1–16.[228] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, Detection of appliance-level abnormal energy consumptionin buildings using autoencoders and micro-moments, in: The Fifth International Conference on Big Data andInternet of Things (BDIoT), 2021, pp. 1–13.[229] Y. Himeur, K. Ghanem, A. Alsalemi, F. Bensaali, A. Amira, Artificial intelligence based anomaly detection ofenergy consumption in buildings: A review, current trends and new perspectives, Applied Energy 287 (2021)116601.[230] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, Smart power consumption abnormality detection in buildingsusing micro-moments and improved k-nearest neighbors, International Journal of Intelligent Systems (2021)1–25.[231] Y. Gong, Z. Jiang, Y. Feng, B. Hu, K. Zhao, Q. Liu, W. Ou, Edgerec: Recommender system on edge in mobiletaobao, in: Proceedings of the 29th ACM International Conference on Information & Knowledge Management,2020, pp. 2477–2484. 34
PREPRINT - F
EBRUARY
16, 2021[232] C. Sun, H. Li, X. Li, J. Wen, Q. Xiong, W. Zhou, Convergence of recommender systems and edge computing:A comprehensive survey, IEEE Access 8 (2020) 47118–47132.[233] X. Su, G. Sperl`ı, V. Moscato, A. Picariello, C. Esposito, C. Choi, An edge intelligence empowered recommendersystem enabling cultural heritage applications, IEEE Transactions on Industrial Informatics 15 (7) (2019) 4266–4275.[234] Y. Himeur, A. Elsalemi, F. Bensaali, A. Amira, The emergence of hybrid edge-cloud computing for energyefficiency in buildings, in: Proceedings of SAI Intelligent Systems Conference, 2021, pp. 1–12.[235] B. Hidasi, Cutting-edge collaborative recommendation algorithms: Deep learning. (2018).[236] A. Felfernig, S. P. Erdeniz, M. Jeran, A. Akcay, P. Azzoni, M. Maiero, C. Doukas, Recommendation technolo-gies for iot edge devices., in: FNC/MobiSPC, 2017, pp. 504–509.[237] A. Sayed, A. Elsalemi, Y. Himeur, F. Bensaali, A. Amira, Endorsing energy efficiency through accurateappliance-level power monitoring, automation and data visualization, in: The 4th International Conferenceon Networking, Information Systems & Security (NISS 2021), 2021, pp. 1–13.[238] X. Wang, B. Gu, Y. Qu, Y. Ren, Y. Xiang, L. Gao, A privacy preserving aggregation scheme for fog-basedrecommender system, in: International Conference on Network and System Security, Springer, 2020, pp. 408–418.[239] X. Zhou, R. Canady, S. Bao, A. Gokhale, Cost-effective hardware accelerator recommendation for edge com-puting, in: 3rd { USENIX }}