Paula Lago

Hybrid Model Rating Prediction with Linked Open Data for Recommender Systems

We detail the solution of team uniandes1 to the ESWC 2014 Linked Open Data-enabled Recommender Systems Challenge Task 1 (rating prediction on a cold start situation). In these situations, there are few ratings per item and user and thus collaborative filtering techniques may not be suitable. In order to be able to use a content-based solution, linked-open data from DBPedia was used to obtain a set of descriptive features for each item. We compare the performance (measured as RMSE) of three models on this cold-start situation: content-based (using min-count sketches), collaborative filtering (SVD++) and rule-based switched hybrid models. Experimental results show that the hybrid system outperforms each of the models that compose it. Since features taken from DBPedia were sparse, we clustered items in order to reduce the dimensionality of the item and user profiles.

A Case Study on the Analysis of Behavior Patterns and Pattern Changes in Smart Environments

Societies need to devise mechanisms of caring for the well aging of the increasing number of seniors, as it is very important for elderly people to maintain their independence. Smart environments are being devised as a form of care in what has been called ambient assisted living. A smart environment should be able to respond in case of emergency or risk and inform any abnormal behavior. Still, not much research is done to understand behavior patterns, temporal changes and other particularities that can affect the effectiveness of smart environments in ambient assisted living. We explored the behavior of two adults in a smart environment in order to reveal temporal, spatial and sequential relations among the activities as well as the changes that these relations undergo overtime and across individuals. This paper presents an analysis of three human behavior patterns: temporal, location and frequency. These patterns are mined on two experimental subjects using the dataset provided by the CASAS project.

The ContextAct@A4H Real-Life Dataset of Daily-Living Activities

Research on context management and activity recognition in smart environments is essential in the development of innovative well adapted services. This paper presents two main contributions. First, we present ContextAct@A4H, a new real-life dataset of daily living activities with rich context data (This research is supported by the Amiqual4Home Innovation Factory, http://amiqual4home.inria.fr funded by the ANR (ANR-11-EQPX-0002)). It is a high quality dataset collected in a smart apartment with a dense but non intrusive sensor infrastructure. Second, we present the experience of using temporal logic and model checking for activity recognition. Temporal logic allows specifying activities as complex events of object usage which can be described at different granularity. It also expresses temporal ordering between events thus palliating a limitation of ontology based activity recognition. The results on using the CADP toolbox for activity recognition in the real life collected data are very good.

Contextualized Behavior Patterns for Ambient Assisted Living

Human behavior learning plays an important role in ambient assisted living since it enables service personalization. Current work in human behavior learning do not consider the context under which a behavior occurs, which hides some behaviors that are frequent only under certain conditions. In this work, we present the notion of a contextualized behavior pattern, which describes a behavior pattern with the context in which it occurs i.e. nap when raining and propose an algorithm for finding these patterns in a data stream. This is our main contribution. These patterns help to better understand the routine of a user in a smart environment, as is evidenced when testing with a public dataset. This algorithm could be used to learn behaviors from users in an ambient assisted living environment in order to send alarms when behavior changes occur.

Learning and managing context enriched behavior patterns in smart homes

Abstract In a society with a growing population of elders, providing efficient and cost-effective long-term care has become central to reducing the economic and societal impact of this demographic shift. Part of the solution to improving the well-being of the elderly at home are smart homes. We examine the role of smart homes in monitoring the activities of the elderly, identifying safety hazards in the home, and understanding environmental changes that may correlate with the deterioration of cognitive and physical health. In this paper, we present LaPlace, a system used to manage context enriched behavior patterns learned in a smart home with sensing devices. LaPlace is based on a formal model to represent intelligible, context enriched behavior patterns, an online adaptive learning algorithm called TIMe which was created for learning such patterns. Context-awareness provides insights about behavior that may otherwise go unnoticed. TIMe is a one-pass algorithm that uses a stream processing model. The TIMe algorithm is presented in this paper along with an extensive evaluation of it using real life datasets.

Representing and Learning Human Behavior Patterns with Contextual Variability

For Smart Environments used for elder care, learning the inhabitant’s behavior patterns is fundamental to detect changes since these can signal health deterioration. A precise model needs to consider variations implied by the fact that human behavior has an stochastic nature and is affected by context conditions. In this paper, we model behavior patterns as usual activity start times. We introduce a Frequent Pattern Mining algorithm to estimate probable start times and their variations due to context conditions using only one single scan of the activity data stream. Experimentation using the Aruba CASAS and the ContextAct@A4H datasets and comparison with a Gaussian Mixture Model show our proposition provides adequate results for smart home environments domains with a lower computational time complexity. This allows the evaluation of behavior variations at different context dimensions and varied granularity levels for each of them.

Context Enriched Patterns of Behavior for Delivering Notifications in Ambient-Assisted Living

Ambient Assisted Living AAL refers to ambient intelligent environments where health and personal care is monitored to help to ensure elders health, safety, and well-being, detecting possible situations that should be watched over. To effectively reason about a current situation and determine its criticality, it should be compared to past patterns. Current systems that learn patterns do not take into account the context in which they occur and we believe adding contextual variables to the description of patterns could improve the reasoning about current situations. Moreover, it can help improve the notification process to dynamically select who and what should be notified, maintaining the privacy of the elder and avoiding spam. In this paper we detail our proposal for a system that adds context to behavior patterns in order to then create personalized messages to the different members of the network of care of an elder living alone.

An Affective Inference Model Based on Facial Expression Analysis

Most of the work in the affective computing field focuses on the recognition of the six basic emotions (sadness, happiness, anger, surprise, disgust and fear). However, when doing academic activities like reading, researching, doing homework or others, these basic emotions are not reported as frequently as others such as interest, confusion and boredom. In this work we propose a computational model of emotion taking into account these three emotions. The model was constructed based on a video analysis of 35 engineering students during two class activities, all of whom reported the emotions they were feeling as they performed the activity. Twenty-two of them were recorded performing a test and the other thirteen were recorded performing a guided activity. From the video, facial expressions were extracted and matched with the emotion reported. The timing of the facial expression was also taken. This allowed us to construct patterns of facial expression and emotion inference rules. Our model is based on distances and indicators of change with respect to a user baseline, which allows the model to adapt to different users, moods and personal manners. Recognizing interest, boredom and confusion can be used as an implicit feedback in recommendation systems.

Mental health information Web search and semantic search extension

MentalWatch is a medical search engine that applies concepts from information retrieval and Web 2.0 to enable patients and doctors to better understand mental illnesses. It puts together scientific articles, news, blog posts and other kinds of contents related to mental health from several Web sources. The main contribution of this work is a software architecture and a semantic search model providing semantic capabilities such as further search suggestions that take advantage of specific domain knowledge. This paper presents an evaluation of the platform taking into account typical information retrieval showing that the results are better than those provided by a general purpose search engine.