Information Filtering via Implicit Trust-based Network
aa r X i v : . [ phy s i c s . d a t a - a n ] D ec Information Filtering via Implicit Trust-basedNetwork
Zhuo-Guo Xuan a , Zhan Li a , and Jian-Guo Liu a , b a Institute of Systems Engineering, Dalian University of Technology, Dalian 116024,People’s Republic of China b Research Center of Complex Systems Science, University of Shanghai for Science andTechnology, Shanghai 200093, PR China
Abstract
Based on the user-item bipartite network, collaborative filtering (CF) recommender sys-tems predict users’ interests according to their history collections, which is a promisingway to solve the information exploration problem. However, CF algorithm encounters coldstart and sparsity problems. The trust-based CF algorithm is implemented by collectingthe users’ trust statements, which is time-consuming and must use users’ private friend-ship information. In this paper, we present a novel measurement to calculate users’ implicittrust-based correlation by taking into account their average ratings, rating ranges, and thenumber of common rated items. By applying the similar idea to the items, a item-basedCF algorithm is constructed. The simulation results on three benchmark data sets show thatthe performances of both user-based and item-based algorithms could be enhanced greatly.Finally, a hybrid algorithm is constructed by integrating the user-based and item-basedalgorithms, the simulation results indicate that hybrid algorithm outperforms the state-of-the-art methods. Specifically, it can not only provide more accurate recommendations, butalso alleviate the cold start problem.
Key words:
Recommender systems, Bipartite networks, Collaborative filtering.
PACS:
Information exploration is one of the results of internet and social network de-velopment. The swift and violent growth of information on the Internet makes itmore and more difficult for users to find available and useful portions [1]. How to
Email address: [email protected] (and Jian-Guo Liu).Preprint submitted toElsevier Science 10 November 2018 elp the users find out the relevant information or products by using the user-itembipartite network is a promising way to solve the information overload problem[2,3,4]. Search engineering and recommender systems are two effective tools tohelp users filter out what pieces are relevant to their tastes. However, search engi-neering presents exactly same list to the same keywords regardless of users’ inter-ests, habits and the history behavior information. Recommender systems filter outthe irrelevant information and recommend the potentially interesting items to thetarget users by analyzing their interests and habits through their history behaviors,which have been successfully applied in a lot of e-commercial web sites [5,6].Collaborative filtering (CF) algorithm is one of the most successful technologies forrecommender systems, which firstly identifies the target user’s neighbors whose in-terests or habits are similar and then presents the recommendation list according tothe neighbor users’ history selections [2,8,9]. Recently, the similar idea has beenapplied to the items. Generally speaking, CF algorithms can be systematically clas-sified as user-based and item-based [1]. User-based methods, regarding each user’sratings as a vector, measure the similarity between the target user and those like-minded people and predict the target user’s rating for the target item according tothe history preferences. User-based CF algorithms have been investigated exten-sively [10]. For example, Herlocker et al. [2] proposed an algorithmic frameworkreferring to user similarity. Luo et al. [12] introduced the local user similarity and global user similarity concepts based on surprisal-based vector similarity and theconcept of maximum distance in graph theory. When the number of items is ap-proximately constant, it is better to give the prediction according to items’ similar-ity network. Item-based methods, regarding each item’s ratings as a vector, measurethe similarity between the target item and other items and predict the target ratingrelying on users’ preferences in history. Because of less updates for average itemsand comparatively static state, the item-based approaches are superior. Sarwar et al. [13] proposed item-based CF algorithm by comparing different items. Deshpande et al. [14] proposed item-based top- N CF algorithm, in which items are rankedaccording to the frequency of appearing in the set of similar items and the top- N ranked items are returned. Recently, Gao et al. [15] incorporated the user rankinginformation into the computation of item similarity to improve the performance ofitem-based CF algorithm.In the previous work, a lot of rating information wasn’t taken into considerationto compute the user or item similarity, such as average ratings, rating ranges, thenumber of users’ common rated items and so on. We argue that, however, theseinformation should be taken into account to measure users’ relationship.When some new users enter into a recommender system, they only give ratings to afew items. Analogously, when some new items are added in the system, they onlyreceive ratings from a few users, which is named cold start problem . It’s very hardto give high quality prediction based on less of history selection information. Inorder to solve the cold start problem, some researchers attempt to integrate user-2ased and item-based CF methods to avoid the limitation of one single algorithm.For instance, Kim et al. [16] built united collaborative error-reflected models thatreflect the average pre-prediction errors of user neighbors and of item neighbors.Jeong et al. [17] proposed an iterative semi-explicit rating method that extrapolatesunrated elements from similar users and items in a semi-supervised manner. Be-sides, Lee et al. [18] used ratings data horizontally and vertically to make two-waycooperative prediction for CF algorithm and thus categorized four possible casesof predictions, namely equivalent case, user-winning case, item-winning case andprediction-impossible case. Empirical experiments show integrating user-based anditem-based methods could enhance the performance greatly.Recently, trust-based mechanism is introduced to alleviate the cold-start problem.Some of e-commerce web sites, such as Epinions, eBay and etc., try to applytrust mechanism to recommend products to consumers. In these web sites, thetrust mechanism is implemented by collecting explicit or implicit trust statements.Explicit trust statements need users to indicate the trust values to their friends[19]. Massa et al. [20] suggested the explicit trust-aware CF recommender sys-tems by searching trust neighbors in depth-first way according to trust propaga-tion. Jamali et al. [19] built a model, named TrustWalker, by random walk in so-cial trust network to find trust neighbors who have rated the target item or similaritems. However, the above trust-based recommendation algorithms need explicittrust statements expressed by users, which are time-consuming and probably ex-pose users’ privacy. Therefore, some implicit trust methodologies are proposed[21,22,23,24,25]. O’Donovan et al. [21] proposed computational models by im-plicit trust based on initial ratings, which only studied the effects of the errors be-tween predicted ratings and actual ratings. Moreover, Kwon et al. [22] created amultidimensional credibility model for neighbor selection in CF algorithm by de-riving source credibility attributes (i.e., expertise, trustworthiness, similarity and at-traction) and extracting each consumer’s importance weight. Li et al. [23] appliedfuzzy logic and inference to support peer recommendation service. Jeong et al. [24] developed user credit-based CF methods which incorporate the information ofeach user’s credit on rating items to compute the aggregation weight. What’s more,Lathia et al. [25] proposed the trusted k -nearest recommenders algorithm whichallows users to learn who and how much to trust others by evaluating the utility ofthe rating information they have received.In previous work, the users’ rating habits wasn’t taken into account, such as averageratings, rating ranges, the number of common rated items and so on. We argue thatthese factors are very important and could be used to measure the implicit trust-based similarities between users or items. In this paper, by constructing the implicittrust-based network, we present three algorithms, say user-based, item-based andhybrid algorithms. The simulation results indicate that these factors are importantand the hybrid algorithm outperforms the state-of-the-art methods and performsvery well to the cold-start problem. 3he following sections are organized as follows: Section 2, we describe the defini-tion and measurement how to calculate the implicit trust-based user or item simi-larity, and the corresponding algorithms are also introduced. In Section 3, the sim-ulation experiments on MovieLens, Netflix and Jester data sets are investigated andthe results are analyzed in detail. Finally, the conclusions are presented and futurework is discussed in Section 4. The meaning of implicit trust-based users can be found in some previous work. Forinstance, O’Donovan et al. [21] supposed that the trustable partners have similartastes and preferences to the target user and they should be trustworthy in the sensethat they have a history of making reliable recommendations, whereas Kwon etal. [22] conceived that trustable neighbors have high expertise, trustworthiness,similarity, etc. In addition, Jeong et al. [24] set the trust-based user as the similarityof voting a rating score with others. Hereinafter,
Trust in recommender system isdefined in the following way. When a user agrees with another user about quality ofcertain products, she probably builds trust relationship with another, which furthermeans that their similar opinions might be inferred in some ways.In this paper, a trust-based user is defined as the user who has the implicit trustrelationship with the target user. Since trust in e-commerce largely depends onsimilar views between users, implicit trust in this paper can be explained as thesimilarity of their opinions and interests on products, which are involved in averageratings, rating ranges and the number of common rated items.
In recommender systems, users express their opinions in the form of reviews, rat-ings, etc. Therefore, we could analyze their interests from different angles to buildcorrespondingly implicit trust among them. In this paper, three factors are takeninto consideration to learn about their interests: 1) users’ average ratings, 2) theranges of their ratings, 3) their common experience. The details are discussed asfollows:1)
Average ratings.
Every user has his/her independent rating schema, i.e. his/her average rating in a4ecommender system, as a result of his/her distinct personal characteristics. Whenusers pay attention to their favorable items, they may express their opinions by vari-ant ratings. In consequence, their independent rating schemas are generated, whichreflect their own characteristics. In traditional CF recommendation algorithms, therating schema is presented as the average rating of a user. For example, lots ofmeasurements are proposed to define the similarities among users, such as Pear-son correlation coefficient [2], adjusted cosine similarity [13], mass diffusion [7,8],heat conduction [3,11] and so on. Empirical studies show that these measurementswith average rating get better results than those without average rating (e.g., co-sine similarity) [26]. In mathematics and statistic domains, average rating reflectsthe general level and the central tendency. Accordingly, in recommender systems,those measurements are used to analyze how far users’ ratings are away from theiraverage ratings and how their ratings evolve. In other words, whatever the ratingsare, if only the differences and extents between users come close, the users are con-sidered similar. In this article, average ratings are taken into account to measure theimplicit trust values between users.2)
Rating ranges .The range of ratings given by a user is probably different for another due to thediversity of users’ habits, mood and contexts. In the practical evaluation, some pes-simistic users under bad mood and contexts fall into the habit of giving low ratingsfor all items. On the contrary, some positive users under good mood and contextsare accustomed to giving high ratings for all items. Since the users do not belong tothe standard-rating sets, they should be treated specifically. Therefore, the range ofratings for every user should be taken into consideration when implicit trust weightis calculated.3)
Common rated items
We suppose the more information we receive from one person, the more we knowabout her. Analogously, in recommender systems, users’ experience is supposed tobe stressed. In recommender systems, the common experience of users that theycontribute to recommendation should be observed in order to improve the perfor-mance of recommender systems. For example, for the target user u and two neigh-bors, say v and w , suppose the similarities between u and w , v are equal, but user u has more common rated items with v than w , therefore, it is reasonable to believethe similarity between u and v is stronger. In our algorithm, common experiencebetween the target user and trustable neighbors is employed entirely.The main principle of implicit trust-based user correlation related to the mentionedthree factors is shown in Fig.1. For user u and v , their implicit trust-based correla-tion is calculated based on their average ratings, say r u and r v , rating ranges, say R max u − R min u and R max v − R min v and the number of their common rated items n .Considering the above three factors, we present the formulation to calculate implicit5 Item h Item i Item k Item j User u
5 4 2 User v
3 4 1 uv min u R max u R min v R max v R u r , u i r v r , v i r , - max min u i uu u r rR R - , max min v i vv v r rR R -- Fig. 1. (Color Online) Implicit trust-based user correlation. For user u and v , their commonexperience is co-rated item i and k . For item i (red rectangle), user u gives normalizeddistance from concrete rating value to average rating within his rating range, and so doesuser v . The definition could be used to measure the complement of absolute differencebetween the two users’ distances combining their common experience. trust between user u and v : C U ( u, v ) = 11 + e − n (cid:18) − n n X i =1 | φ ( u ) − φ ( v ) | (cid:19) (1)where φ ( u ) = r ui − r u R max u − R min u and n is the number of common rated items for user u and v . The sigmoid function, / (1 + e − n ) , is used to rectify weight by the numberof common rated items, n , which has ever been distinctly used to adjust PearsonCorrelation coefficient [19]. In this paper, K -nearest neighbors of the target user are evaluated to investigate theeffect of implicit trust-based correlation on cold start problem. Afterwards, the pre-dicted rating, b r Uuj from user u to the target item j is given according to the followingformulation. b r Uuj = r u + P v ∈ Γ u C U ( u, v )( r vj − r v ) P v ∈ Γ u C U ( u, v ) (2)where Γ u is a set of the nearest neighbors of user u , and C U ( u, v ) is the implicittrust-based correlation between user u and v obtained by Eq.(1). Introducing a similar idea on the item correlation definition, the effect of the im-plicit trust-based correlation on item-based CF algorithm is investigated.6 .2.1 Definition of implicit trust-based items
When we are satisfied with products that we have purchased, we usually place themin trusted zone. Perhaps, in the future, we will buy them again. On the contrary, ifwe complain about some bad products, we place them in restricted zone and wemay never buy them again.In this paper, items based on implicit trusts are considered relying on proximityof items that a user has evaluated in her history. From this point of view, trusteditems can be explained as the items that are close to those that one user trusts.In other words, while a user set a certain item in her trusted zone, the trust-baseditems, in terms of intrinsic attributes, accepted degrees, rating values and commonpopularity, are very similar to it. The process to search implicit trust-based items isto analyze all users’ opinions about these items.
In the paper, like implicit trust-based user correlation definition, three factors arereferred to compute items’ implicit trust-based relationship, which can be describedas: 1) the internal or intrinsic attributes of an item, 2) the accepted degree of an item,3) the common rated times between any pairs of items. The detail is described asfollows:1)
Intrinsic attributes of an item
The internal attributes of an item determine all users’ opinions about it. In otherwords, the average rating reflects intrinsic attributes of the item. If the quality ofan item is good, users generally like it and give it high ratings, and vice versa. Themore users have evaluated an item, the closer the average rating is to the internalcharacteristics of the item. The average rating implies all users’ opinions about theitem.We primarily pay attention to the distance from concrete rating to average rating.That means, the nearer to average rating the concrete rating value is, the more trust-worthy an item is. In a word, average rating plays a significant role to implementrecommendation based on implicit trust-based items.2)
Accepted degree of an item
The accepted degree of an item can be observed from two perspectives, minimumrating and maximum rating, which can be inferred from the rated range of the item.To an item, the minimum rating shows how bad the item a user thinks and themaximum one shows how good the item she considers. In brief, the minimum andmaximum ratings describe the accepted degree of an item derived from all users’opinions. For instance, if a movie is rated with low ratings.7)
Common rated times between items
The number of users who have commonly rated the items could affect the trustwor-thy levels of the items.The more users give high ratings to two items, the more correlated these itemsare. Generally, the number of common rated times between the target item and itsimplicit trust-based neighborhood items should be taken into account.The core principle of implicit trust-based item correlation is depicted in Fig.2. Foritem i and j , the intrinsic attributes are denoted as their average ratings r i and r j respectively. The differences between maximum and minimum ratings are denotedas R max i − R min i and R max j − R min j . The number of common rated times is set as m . max i R min i R i r , v i r , max min v i ii i r rR R -- i max j R min j R j r , v j r , -- max min v j jj j r rR R j Item j Item i User t
5 User v
4 3 User w
2 4 User u Fig. 2. (Color Online) Three factors affecting implicit trust weight of credible items. Foritem i and j , their common popularity is the aggregation of over v ’s and w ’s ratings. Foruser v (red rectangle), item j gets normalized distance from concrete rating value to averagerating within its rating range, and so does item i . The goal is to aggregate the complement ofabsolute difference between the two items’ distances combining their common popularity. Therefore, the following formulation could be given, C I ( i, j ) = 11 + e − m (cid:18) − m m X v =1 | φ ( i ) − φ ( j ) | (cid:19) (3)where φ ( i ) = r vi − r i R max i − R min i and m denotes the number of users who have rated bothitem i and j . The sigmoid function, / (1 + e − m ) , is used to rectify weight bycommon users. To investigate the effect of implicit trust-based item correlation network on users’cold start problem, the K -nearest neighbors are evaluated in this paper. The pre-dicted rating from user u to item j is given according to the following item-basedCF algorithm. b r Iuj = r j + P i ∈ Γ j C I ( i, j )( r ui − r i ) P i ∈ Γ j C I ( i, j ) (4)8here Γ j is a set of the nearest neighbors of item j , and C I ( i, j ) denotes the implicittrust weight between item i and j by Eq.(3). Traditional CF algorithm encounters cold start problem because of data set sparsity,which can be further divided into cold start users and cold start items [16]. A coldstart user indicates the new user who has participated in recommendation but hasexpressed few opinions. In this situation, it is often the case that there is no inter-section at all between two users, and it is difficult to calculate the user similaritybased on common rated items. Even when the computation of similarity is possible,it may not be very reliable because of the insufficient information available. A coldstart item is caused by the new item. In the CF-based recommender systems, thisitem cannot be recommended due to insufficient user opinions. The simulation re-sults indicate that the hybrid algorithm could not only greatly enhance the accuracy,but also effectively solve the cold start problem.In this paper, to alleviate the cold start problem, we present a hybrid recommen-dation algorithm by integrating implicit trust user-based and item-based CF algo-rithms, where the predicted rating is given in the following way b r uj = (1 − α ) b r Uuj + α b r Iuj , (5)where b r Uuj is the prediction rating based on user-based CF algorithm in Eq.(2), b r Iuj isthe prediction rating based on item-based CF algorithm in Eq.(4), and α is a tunableparameter whose range is [0,1]. When α = 0 , the hybrid algorithm degenerates tothe user-based algorithm, and it becomes the item-based CF algorithm when α = 1 .We can adjust value to control the ratios from the above two algorithms and findthe optimum solution. In this paper, our simulation experimental data comes from MovieLens , Netflix and Jester. The Movielens data is collected by the GroupLens Research Project dur-ing the seven-month period from September 19th, 1997 through April 22nd, 1998.The dataset consists of 100,000 ratings from 943 users on 1,682 movies and rating able 1Basic statistics of the test data sets.Data Sets Users Objects Links SparsityMovieLens 6,040 3,592 750,000 . × − Netflix 10,000 6,000 701,947 . × − Jester 2,350 100 169,655 0.7219 scale is from 1 (awful) to 5 (must see), which has been cleaned up so that userswho had less than 20 ratings or did not have complete demographic informationwere removed from this dataset. The Netflix and Jester data are random samplesof the whole records of user activities in Netflix.com and Jester, in which the Net-flix data consists of 10000 users, 6000 movies and 824802 links, and the Jesterdata has 2350 users, 100 jokes and 169,655 connections. Table gives the statisticalproperties of the test data sets.
In order to measure the performances of the present algorithms, the mean absoluteerror (MAE) [27], the root mean square error (RMSE) [19] and the hit rate (HR)are used.
MAE is the mean absolute difference between an actual and a predicted ratingvalue, which is generally used for the statistical accuracy measurements in variousalgorithms. The smaller MAE an algorithm achieves, the better the experimentalresult is. The metric MAE is defined as:
MAE = P n r i =1 | b r i − r i | n r (6)where b r i and r i represent the predicted and actual rating respectively, and n r de-notes the number of tested ratings. RMSE has been typically used to measure the large errors in extreme cases. Anal-ogously, the smaller the value of RMSE an algorithm obtains, the more precise therecommendation is. The metric RMSE is usually defined as follows
RMSE = s P n r i =1 ( b r i − r i ) n r (7)10 .2.3 Hit Rate The hit rate (HR) is also introduced to measure the accuracy of the recommenda-tion. Here, HR is defined as the ratio of the number of hits (i.e., the fraction ofthe number of recommended items and actually chosen items) to the size of therecommendation list. In the information retrieval literature, it is usually equivalentto the metrics
Precision and
Recall . The bigger the value of HR is, the better analgorithm. Formally, HR is defined as HR = HL , (8)where L is the length of recommendation list and H is the percentage of items inthe test set existing in the top- L positions of recommendation list. The implicit trust-based effects are implemented on user-based, item-based andhybrid algorithms separately. Since the prediction performance is influenced by thesize of the K nearest neighbors, it is essential to determine a proper size of thenearest neighbors Top K , where K is set as 3, 5, 10, 15 and 20 respectively. Sincethe typical length for recommendation list is ten items, our experiments set L =10.The parameter α is adjusted in the interval [0, 1] and the increment is 0.1. In this section, we investigate the performance of the user-based CF algorithm (de-noted as IU-CF) and compare it against the performances of classic user-based CFusing well-known Pearson Correlation coefficient (denoted as PCF) and adjustedcosine-based CF algorithm [13] (denoted as AC-CF).Figure 3 illustrates the results of MAE, RMSE and HR for PCF, AC-CF, IU-CFand II-CF algorithms respectively. The results demonstrate that IU-CF and II-CFalgorithms enhance the performance of the initial two approaches, PCF and AC-CF.From Fig. 3, one can see that MAE of IU-CF algorithm has the lowest level in thethree algorithms. As the number of the nearest neighbors K increases, the MAEcurves of all four algorithms tend to decrease, which implies that more neighborscan make better prediction although computation and time complexity is high. TheRMSE results in Fig.3 show that IU- and II-CF algorithms have the smallest errorsin the three algorithms while PCF algorithm gets results with the largest errors.In other words, our approach can predict more accurately than PCF and AC-CFalgorithms. In addition, the similar RMSE downtrend for all algorithms appears inFig.3 as the growth of the sizes of user neighborhood. Fig.3 illustrates the resultsof HR of three algorithms. As shown in Fig.3, at most neighborhood sizes, HR of11 ig. 3. (Color Online) Comparison of results achieved by Pearson-based (PCF), adjustedcosine-based (AC-CF), implicit trust-based (IU-CF) and item-based CF (II-CF) algorithms.Note that both IU- and II-CF have the smallest MAE, RMSE and highest HR for Movielens,Netflix and Jester data sets. IU- and II-CF algorithms are remarkably better than the results of PCF and AC-CFalgorithms. Even though only a minority of neighbors participate in prediction, thepresent IU- and II-CF algorithms outperform the other two methods. And, when thenumber of nearest neighbors increases, the curves of the three methods graduallychange upward and finally tend to become flat. From the results of Fig.3, it canbe concluded that the present user-based and item-based approaches can providebetter recommendations.
In this section, the effects of the implicit trust-based correlations on hybrid rec-ommendation (HCF) are investigated by integrating the user-based and item-basedCF algorithm. In the experiment, we compare hybrid recommendation against theabove two pure algorithms with different values. Figure 4 summarizes the exper-iment results of MAE, RMSE and HR for HCF algorithm according to the value α variation. We examine the HCF results of the three metrics in order to choose12 .991.021.051.080.991.021.05 Netflix R M SE / R M SE m i n M A E / M A E m i n M A E / M A E m i n R M SE / R M SE m i n HR / HR m i n HR / HR m i n K = 3 K = 5 K = 10 K = 15 K = 20
Movielens
Fig. 4. (Color Online) Comparison of setting different α values for HCF. Note that MAEand RMSE first fall down before α =0.5 and then climb up after that for Movielens andNetflix data sets. HR obeys reverse distribution with the boundary α =0.5. The conclusionis that the optimal α is 0.5 for HCF. optimal parameter α . In the experiment, the value is continuously changed in theinterval [0, 1] with the increment 0.1. From Fig. 4, MAE and RMSE apparently de-crease as the value of increases from 0 to 0.5; after this point 0.5, the upward MAEand RMSE gradually appear for Movielens and Netflix data sets. On the contrary,the metric HR considerably ascends before the value 0.5 and after that it begins todescend steadily. The optimal parameter for Jester data set is not exactly 0.5, butalso close to this value. The results indicate that the optimal value is 0.5 no matterwhich metric is evaluated for HCF.Figure 5 illustrates the comparison of IU-CF, II-CF and HCF in the metrics MAE,RMSE, and HR respectively at the increasing sizes of the neighborhood from 3to 20 when the optimal parameter α is 0.5. As shown in the Fig.5, for Movielens,Netflix and Jester data sets, HCF obtains the remarkably lowest levels of MAE andRMSE in the three methods when K is quite small, as well as highest HR val-ues. Summing up the above three metric results, the conclusion can reasonably bedrawn that HCF which integrates recommendations by implicit trust-based user anditem similarity network can further improve the performance of recommendationin some degree than pure IU-F and II-CF. More importantly, HCF could efficientlysolve cold start problem. 13 ig. 5. (Color Online) Comparison of results achieved by IU-CF, II-CF and hybrid CF(IU+II CF) algorithms algorithms. Note that hybrid CF algorithm has the smallest errorsand highest hits. Information is explored dramatically in the social network era. According to thestructural properties of web connections, search engineering could help us to digout the most relevant web page according to the keywords. However, search engi-neering couldn’t help users find the fresh information or products related to theirinterests and habits, and couldn’t analyze their personation, either. Based on theuser-item bipartite network, recommender system is a promising tool to dig out thevaluable information for the users. However, the existing user or item correlationdefinition didn’t take into account the users’ rating habits and statistical proper-ties in detail. Traditional CF algorithm suffers the cold-start problem, and explicittrust-based recommender systems require users to express explicit trust statements,which may be time-consuming and expose privacy of users. Besides, the existingimplicit trust-based algorithms take few factors into consideration to calculate thetrust weight. Therefore, their recommendation results are not sufficiently accurate.This work addresses these problems by introducing implicit trust-based correlationnetwork. When computing implicit trust weight, we fully consider implicit trust-14ased factors about users (e.g. average ratings, rating ranges, and common experi-ence) and items (e.g. internal attributes, accept degrees and common popularity).The simulation results show that the proposed implicit user-based, item-based andhybrid algorithms solve cold start problem and provide accurate recommendations.Although our approaches presented in this article have shown encouraging results,we also have several interesting tasks for future work. First, we are going to focuson doing research on transitive trust. In this paper, we have just paid attention tocomputing the implicit trust weight, but have not studied trust propagation. In realsocial network, trust can propagate from one person to another. Due to trust propa-gation, perfect neighbors are easy to be accessed and the cold start problem couldalso be overcome in some degree. In the future, we are going to take transitivetrust into consideration in order to improve the performance of implicit trust-basedrecommender system. Second, we attempt to append robust mechanisms againstthe attacks by malicious users to improve our proposed approaches. The reason isthat some e-commerce online recommender systems at present are often attackedby negative canvassers. Therefore, it is worthwhile to emphasize the robustness ofan algorithm as an important aspect of practical recommender systems. Finally, weplan to develop new evaluation metrics to assess the performance of trust-basedalgorithms because the current metrics seldom examine the robustness of recom-mender systems.
Acknowledgments
We acknowledge the GroupLens Research Group for MovieLens data. This workhas been partly supported by the Natural Science Foundation of China (Grant Nos.10905052, 71171136, 71031002), the Fundamental Research Funds for the CentralUniversities of China under Grant No. DUT11RW422, and the Innovation Pro-gram of Shanghai Municipal Education Commission (11ZZ135, 11YZ110). JGL issupported by the Shanghai Leading Discipline Project (S30501), Shanghai Rising-Star Program (11QA1404500) and Key Project of Chinese Ministry of Education(211057).
References [1] G. Adomavicius, A. Tuzhilin. IEEE Tran. on Know. & Data Eng. 17 (2005) 734.[2] J.L. Herlocker, J.A. Konstan, A. Borchers, J. Riedl. In Proc. of SIGIR ’99, 1999, pp.230.[3] T. Zhou, Z. Kuscsik, J.-G. Liu, M. Medo, J.R. Wakeling, Y.C. Zhang. Proc. Natl. Acad.Sci. U.S.A. 107 (2010) 4511.
4] T. Zhou, R.-Q. Su, R.-R. Liu, L.-L. Jiang, B.-H. Wang, Y.-C. Zhang. New J. Phys. 11(2009) 123008.[5] G. Linden, B. Smith, J. York. IEEE Internet Computing 7 (2003) 76.[6] L. Lu, Y.-C. Zhang, C.-H. Yeung, T. Zhou. PLoS One 6 (2011) e21202.[7] T. Zhou, J. Ren, M. Medo, Y.-C. Zhang, Phys. Rev. E 76 (2007) 046115.[8] J.-G. Liu, B.-H. Wang, Q. Guo. Int. J. Mod. Phys. C 20 (2009) 285.[9] J.-G. Liu, T. Zhou, H.-A. Che, B.-H. Wang, Y.-C. Zhang. Physica A 389 (2010) 881.[10] D. Sun, T. Zhou, J.-G. Liu, R.-R. Liu, C.-X. Jia, B.-H. Wang. Phys. Rev. E 80 (2009)017101.[11] J.-G. Liu, T. Zhou, Q. Guo. Phys. Rev. E 87 (2011) 037101.[12] H. Luo, C. Niu, R. Shen, C. Ullrich. Machine Learning 72 (2008) 231.[13] B. Sarwar, G. Karypis, J. Konstan, J. Riedl. In Proc. of ACM on WWW conference2001, pp. 285.[14] M. Deshpande, G. Karypis. ACM Tran. Info. Sys. 22 (2004) 143.[15] M. Gao, Z.-F. Wu, F. Jiang. Info. Proc. Lett. 111 (2011) 440.[16] H.N. Kim, A. El-Saddik, G.S Jo. Decision Support Systems 51 (2011) 519.[17] B. Jeong, J. Lee, H. Cho. Expert Systems with Applicatios 36 (2009) 6181.[18] J.S Lee, S. Olafsson. Expert Systems with Applicatios 36 (2009) 5353.[19] M. Jamali, M. Ester. In Proc. of the 15th ACM SIGKDD, 2009, pp. 397.[20] P. Massa, P. Avesani. In Proc. of Federated International Conference on the Move toMeaningful Internet, 2004, pp. 492.[21] J. O’Donovan, B. Smyth. In Proc. of the 10th International Conference on IntelligentUser Interfaces, 2005, pp. 167.[22] K. Kwon, J. Cho, Y. Park. Expert Systems with Applicatios 36 (2009) 7114.[23] Y.M. Li, C.P. Kao. Expert Systems with Applicatios 36 (2009) 3263.[24] B. Jeong, J. Lee, H. Cho. Expert Systems with Applicatios 36 (2009) 7309.[25] N. Lathia, S. Hailes, L. Capra. In Proc. of IFIPTM 2008:Joint iTrust and PSTconferences on Privacy, Trust management and Security, 2008, pp.119.[26] J.S. Breese, D. Heckerman, C. Kadie. In Proc. Of the 14th Conference on Uncertaintyin Artificial Intelligence, 1998, pp.43.[27] J.L. Herlocker, J.A. Konstan, L.G. Terveen, J. Riedl. ACM Transactions onInformation Systems 22 (2004) 5.4] T. Zhou, R.-Q. Su, R.-R. Liu, L.-L. Jiang, B.-H. Wang, Y.-C. Zhang. New J. Phys. 11(2009) 123008.[5] G. Linden, B. Smith, J. York. IEEE Internet Computing 7 (2003) 76.[6] L. Lu, Y.-C. Zhang, C.-H. Yeung, T. Zhou. PLoS One 6 (2011) e21202.[7] T. Zhou, J. Ren, M. Medo, Y.-C. Zhang, Phys. Rev. E 76 (2007) 046115.[8] J.-G. Liu, B.-H. Wang, Q. Guo. Int. J. Mod. Phys. C 20 (2009) 285.[9] J.-G. Liu, T. Zhou, H.-A. Che, B.-H. Wang, Y.-C. Zhang. Physica A 389 (2010) 881.[10] D. Sun, T. Zhou, J.-G. Liu, R.-R. Liu, C.-X. Jia, B.-H. Wang. Phys. Rev. E 80 (2009)017101.[11] J.-G. Liu, T. Zhou, Q. Guo. Phys. Rev. E 87 (2011) 037101.[12] H. Luo, C. Niu, R. Shen, C. Ullrich. Machine Learning 72 (2008) 231.[13] B. Sarwar, G. Karypis, J. Konstan, J. Riedl. In Proc. of ACM on WWW conference2001, pp. 285.[14] M. Deshpande, G. Karypis. ACM Tran. Info. Sys. 22 (2004) 143.[15] M. Gao, Z.-F. Wu, F. Jiang. Info. Proc. Lett. 111 (2011) 440.[16] H.N. Kim, A. El-Saddik, G.S Jo. Decision Support Systems 51 (2011) 519.[17] B. Jeong, J. Lee, H. Cho. Expert Systems with Applicatios 36 (2009) 6181.[18] J.S Lee, S. Olafsson. Expert Systems with Applicatios 36 (2009) 5353.[19] M. Jamali, M. Ester. In Proc. of the 15th ACM SIGKDD, 2009, pp. 397.[20] P. Massa, P. Avesani. In Proc. of Federated International Conference on the Move toMeaningful Internet, 2004, pp. 492.[21] J. O’Donovan, B. Smyth. In Proc. of the 10th International Conference on IntelligentUser Interfaces, 2005, pp. 167.[22] K. Kwon, J. Cho, Y. Park. Expert Systems with Applicatios 36 (2009) 7114.[23] Y.M. Li, C.P. Kao. Expert Systems with Applicatios 36 (2009) 3263.[24] B. Jeong, J. Lee, H. Cho. Expert Systems with Applicatios 36 (2009) 7309.[25] N. Lathia, S. Hailes, L. Capra. In Proc. of IFIPTM 2008:Joint iTrust and PSTconferences on Privacy, Trust management and Security, 2008, pp.119.[26] J.S. Breese, D. Heckerman, C. Kadie. In Proc. Of the 14th Conference on Uncertaintyin Artificial Intelligence, 1998, pp.43.[27] J.L. Herlocker, J.A. Konstan, L.G. Terveen, J. Riedl. ACM Transactions onInformation Systems 22 (2004) 5.