Collaborative Filtering (CF) is the most popular recommendation technique. ... exploits social networks techniques, such as transitivity, to explore additional links ...
From Social Networks to Behavioral Networks in Recommender Systems Ilham Esslimani, Armelle Brun, Anne Boyer KIWI Team, Universit´e Nancy 2, LORIA 615 rue du Jardin Botanique, 54600 Villers-L`es-Nancy, France {ilham.esslimani, armelle.brun, anne.boyer}@loria.fr Abstract—Recommender systems are widely used for personalization of information on the web and information retrieval systems. Collaborative Filtering (CF) is the most popular recommendation technique. However, classical CF systems use only direct links and common features to model relationships between users. This paper presents a new Collaborative Filtering approach (BNCF) based on a behavioral network that uses navigational patterns to model relationships between users and exploits social networks techniques, such as transitivity, to explore additional links throughout the behavioral network. The final aim consists in involving these new links in prediction generation, to improve recommendations quality. BNCF is evaluated in terms of accuracy on a real usage dataset. The experimentation shows the benefit of exploiting new links to compute predictions. Indeed, BNCF highly improves the accuracy of predictions, especially in terms of HMAE.
I. I NTRODUCTION Social networks represent a social structure between actors, mostly individuals or organizations. It indicates the ways they are connected through various social relationships as friendship, co-working or information exchange [1]. With the evolution of the web, social network analysis is becoming increasingly relevant, since it aims at understanding the evolution of interactions between users and social flows. The development of the web engendered also an important proliferation of information resources. The need of automatic personalization of information thus becomes heightened. Recommender systems are widely used for this purpose thanks to their ability to analyze users behaviors and guide them towards relevant resources that suit their preferences. Recommender systems use different input data to construct user models in order to generate recommendations. The input data can include content informations [2], explicit data like votes [3], demographic data [4], etc. Collaborative Filtering (CF) is one recommendation technique that identifies relationships (similarities) between users, based on their ratings in order to select neighbors and compute predictions for the active users. Despite the success of recommender systems and collaborative filtering in many application areas, some research questions still remain. Some of these questions concern the requirement of explicit rating data to compute similarities between users. As explicit rating data is not always available, one challenge for recommender systems is to take into consideration another type of data that represent efficiently users
behaviors. In this context, usage traces can be a relevant source of data. Another challenge for recommender systems is to model relationships between users, not by using rating information or social links, but navigational behaviors instead. Additionally, in terms of modeling, one research problem consists in identifying new links between users that are not directly connected. Indeed, classical predictive systems exploit only direct links and common preferences to compute recommendations. In this frame, so as to improve the quality of predictions, we propose to use social networks techniques, especially transitive associations to explore new links. New “neighbors” can be then used to compute predictions. Thus, the research problem we are interested in, is related to two main points: •
•
By using usage traces, how can we evaluate correlations between users and how can we construct behavioral networks. How can we introduce social networks techniques in order to model relationships between users and identify new links in these behavioral networks.
In this paper, we explore these issues and propose a promising model that we experiment on a real usage dataset. We suggest a new Behavioral Network based Collaborative Filtering system (BNCF), that exploits navigational patterns and transitive links to model users. To analyze behavior similarities, we employ navigational patterns by taking into account the longest common sub-sequences between pairs of users. Then, users are modeled through a behavioral network based on these navigational similarities. In social networks, transitivity means “the friends of my friends are my friends”. The transposition of this property in behavioral networks implies “the users looking like those who look like me, look like me”. The application of social networks based techniques such as transitivity aims at refining the behavioral network by additional links that can enhance the performance of the CF system. This paper is organized as follows. We describe in the second part some research studies related to the analysis of usage traces and graph based recommender systems. In the third part of the paper, we present the BNCF approach that exploits social networks techniques. The fourth part describes the experimentation. Then, the results of the experimentation
are put forward in the fifth part and finally we present a conclusion. II. R ELATED WORK
The common feature of all these approaches related to WUM, is the exploitation of usage traces so as to compute links, distances and relationships between users based on common visited resources in order to generate predictions.
A. Analysis of Usage traces Several studies describe the impact of usage traces on the recommendation process in predictive and recommender systems. These studies demonstrate how the analysis of navigational activities can be a relevant method that allows to model users behaviors and identify their potential needs [6]. Let us notice that analysis of usage traces is mainly related to the area of Web Usage Mining (WUM) which aims at observing users behaviors while interacting with a system. This observation refers to direct traces as explicit ratings and annotations, or non direct traces like bookmarking, frequencies of visits, visited links, etc. from which users preferences can be inferred. The relevance of considering usage traces in order to learn implicit votes is presented in [7]. A metric is suggested to compute the “page interest estimator” from non direct traces. Frequent patterns mining, Longest Common Subsequences (LCS) technique and Markov models, are some of the WUM approaches that tend to harness the navigational activities in order to analyze users behaviors. The attempt of frequent patterns mining is the discovering of time ordered sequences that have been followed by past users in order to predict future resources [8]. In order to discover patterns from these traces, the process consists, first of all, in usage data preprocessing [9], then it runs the patterns discovery mechanism which allows finally the generation of recommendations. The recommended resources are represented as pages or resources that are frequently accessed by related users (with common preferences). In [10] the concept of “active session window” is used to generate recommendations. The recommendation system suggests a hybrid personalization model that switches between different recommendation models according to the degree of connectivity and the depth of the active user’s session. Discovering of Longest Common Subsequences (LCS) is another technique that has been applied in WUM domain in order to analyze the potential links between navigational paths and users profiles. Basically, this technique is one dynamic programming method, it aims at identifying the longest common subsequence relating to two given sequences. An LCS based architecture is suggested in [11] for classifying navigational patterns and generating predictions to users. In [12] an algorithm based on LCS technique is proposed for clustering users by using their navigational data. This clustering approach uses the similarities between two navigational paths based on the LCS and the time spent on resources contained in this LCS. Another approach that uses sequential links for navigational activities is Markov chain model. In accordance with [13], the sequential dependencies of navigational behaviors of users are modeled by Markov Models ; the conditional probability of one resource, considering users navigational traces is computed.
B. Graph based Recommender Systems With the heightened development of the web, social networks have been subject of several research studies. Most of them study and analyze social networks structures to represent interactions, collaborations or influences between entities. [14] present a measurement analysis of various online social networks by studying the corresponding topological properties. [15] suggest a mixture between usage traces and social networks. Interactions between social influence and similarity of activities are modeled in order to predict future behaviors. An important attention has been also devoted to the combination of social networks and recommender systems. Some studies like [16] incorporate social network information into Collaborative Filtering. Nodes are consumers and links are the social relationships among them. In order to identify neighbors, distances between users in the social network are harnessed. These distances are computed by the breadth-first search and are used to calculate predictions. However, most of the studies in the context of recommender systems, do not use directly social information, but apply rather graph based techniques to model users in order to make predictions. A navigation graph based recommender system is presented in [17]. A distance metric based on maximum common sub-graph is used to compute distances between nodes that represent navigational sequences. Then, these nodes are clustered by using the computed distances and the recommender system matches the active user to the nearest cluster and suggest recommendations according to his navigational sequence. A two layer graph based recommender system that combines content and collaborative approaches in the context of digital libraries is proposed in [18]. Books and users represent layer nodes and the edges represent transactions. Low-degree association (based on content and collaborative similarity weights) and high-degree association graph (based on Hopfield spreading activation algorithm) have been applied in order to improve recommendations quality. Transitive relationships related to associative retrieval approaches, have been explored in the context of Collaborative Filtering in order to alleviate the sparsity problem [19]. The transitive associations among users are identified based on spreading activation techniques. III. B EHAVIORAL N ETWORK BASED CF (BNCF) In standard recommender systems, similarities between users are evaluated based on common features as ratings, visited resources, etc. However, if two users do not share any of these features, we cannot establish links between them. Thus, to overcome this problem we propose to use social networks techniques such as transitivity in order to identify new links between users. The finality consists in involving
these new links in the recommendation process in order to improve predictions accuracy. We propose the construction of a behavioral network by modeling links between users, based on their behavioral similarities that are computed by using navigational patterns. Unlike classical predictive systems based on navigational patterns (presented in section II-A), we attempt to analyze behavioral similarities between pairs of users by using usage traces. Then, by applying transitivity, we aim at identifying new neighbors that are strongly connected to the active user throughout the network. Strong connections are then deduced from high similarity values between the intermediate neighbors. The following sections present in details the different mechanisms used by the BNCF system to generate predictions. A. Construction of the Behavioral Network In social networks approaches, graph based models are employed to model networks structures based on social information. In this paper, we introduce another type of network, based on behavioral information where users are not necessarily connected in the real world as social networks, but are linked as they share similar navigational patterns. The following section presents the technique that we employ to assess navigational similarities between users in order to construct the behavioral network. 1) Computing navigational similarities: As presented in [20], we consider that two users ua and ub , who share common sequential patterns are highly similar. The longer the sequence of a common pattern is, the more the users are similar. Therefore, our goal is to identify for every pair of users < ua , ub >, the maximum length LKmax (ua , ub ) of a pattern among their common patterns. Then, the similarity of navigation between two users is computed by using Equation 1 that takes into account the following parameters: • Common patterns between the active user ua and the neighbor user ub . • The maximum length of a common pattern between the active user ua and the neighbor user ub . • The maximum length of sessions. This formula computes, for each pair of users ua and ub the similarity of navigation SimN av(ua ,ub ) as the ratio of the maximum length of a common frequent pattern LKmax (ua , ub ) and the minimum of maximum sizes of ua and ub sessions denoted SessM ax(ua ) and SessM ax(ub ) . We note that the common frequent pattern is intra-session. SimN av(ua ,ub ) =
LKmax (ua , ub ) (1) min(SessM ax(ua ) , SessM ax(ub ) )
We use the minimum of maximum sizes of sessions in the denominator so as to avoid to penalize a new user who has few sessions with short sizes. We note that the correlation value is normalized between 0 and 1. This metric emphasizes the importance of the longest frequent patterns to evaluate similarities of users. The higher the length of a sequential pattern is, the more the users are similar.
Fig. 1.
Identification of new neighbors in the Behavioral Network
2) Modeling the Behavioral Network: In order to model links between users, we use a directed graph G = (V, E) where vertices V represent users, edges E represent the links between users and the navigational similarities (computed in the previous step) represent the weights of the edges. We employ the Floyd-Warshall algorithm in view of its efficiency and its simple implementation. This algorithm is used to compute the shortest path between every pair of nodes ua and ub in the graph by taking into account the weights. The algorithm checks whether a shorter path from nodes ua to ub exists via nodes ur . At the end of the process, the matrix contains the length of the shortest path from ua to ub . To use this algorithm, we transform the navigational similarities into distance-like values. The computation of the shortest paths leads to the identification of new links between users throughout transitive relationships, as described in Figure 1. Indeed, as we can see, initially the active user ua has two direct neighbors ue and ub . But, with the transitive links, he can be connected to additional neighbors like user ud from two possible paths. Thus, the algorithm selects the shortest path. That means that even if two users are not similar in terms of navigation (they have not viewed the same resources in the past) and are not directly connected in the behavioral network, we can find strong links between them thanks to the strong similarities of intermediate neighbors. The smaller the distance is, the more these two users are similar. This step allows the discovering of new potential neighbors of active users that are not direct neighbors. These neighbors are then involved in the stage of computing predictions with the objective of improving the performance of the recommender system. B. Prediction generation 1) Estimating ratings from usage traces: As ratings are required in the prediction step, we employ usage traces to estimate them as mentioned in section II-A. We choose two implicit parameters: frequencies of visiting a resource and duration of visiting a resource. Considering an active user ua , the frequency of visiting an item ik is the ratio of the number of visits of ik (N(ua ,ik ) ) and the average number of visits on all items I (N(ua ,I) ) as described in Equation (2).
F requency(ua ,ik ) =
N(ua ,ik ) N(ua ,I)
(2)
As regards duration, it is computed as the ratio of the duration of visiting an item ik (Drt(ua ,ik ) ) and the total duration of visiting all items I (Drt(ua ,I) ) as presented in Equation (3). Duration(ua ,ik )
Drt(ua ,ik ) = Drt(ua ,I)
(3)
Once frequencies and durations are calculated, we use the formula 4 suggested by [21], in order to compute and normalize our ratings according to the rating scale. P
p(c) ∗ c(ua , ik ) Vmax − Vmin P ∗ ) cmax c p(c) (4) fT ransf(ua ,ik ) represents the transformation function of ua rating on item ik . Vmin and Vmax are respectively the minimum and maximum possible ratings according to the rating scale. p(c) denotes the weight assigned to the criterion (Frequency and Duration in our case), c(ua ,ik ) is the value of the criterion and cmax denotes the maximum value of the criterion. 2) Computing predictions: Once the links (shortest paths) between the active user and other users are computed, the similarities are deduced from normalized distance-like values. Then, predictions are calculated by using the weighted average prediction formula used by CF [22]. We select nearest neighbors ub (direct and non direct) in the behavioral network, that have already rated the active item ik . c
fT ransf(ua ,ik ) = Vmin +(
P
Sim(ua ,ub ) ∗ (V(ub ,ik ) − V(ub ) ) P ub Sim(ua ,ub ) (5) Here, P red(ua ,ik ) represents the prediction of the rating of the active user on ik and Sim(ua ,ub ) the similarity value between ua and ub based on the shortest path between them. Items that will be recommended to the active user are the ones with high predicted rating values. P red(ua ,ik ) = V(ua ) +
ub
IV. E XPERIMENTATION A. Datasets In order to evaluate the performance of our CF system, we use real usage datasets extracted from the intranet of Credit Agricole Banking Group, in particular the usage data relating to the Department of Strategies and Technology Watch. All the users are members of the Group and can access various informations like: news, articles, faq, special reports, etc. This intranet contains numerous resources and web pages. Therefore, the finality of integrating a recommender system consists in guiding users towards relevant resources corresponding to their profiles. Thus, to train our model, we use the usage data that reflects the navigational activities of users. This data has been collected during 24 months and stored in server log files. It contains mainly informations about anonymous user-ids,
session-ids and time of starting sessions. The selected dataset is related to 748 users and 3856 resources. It has been split into 80% and 20% corresponding respectively to training and test datasets by taking into account the temporal dimension. We fixed the rating scale to [1 − 5] for estimated ratings. B. Evaluation Different evaluation metrics can be used in the experimentation of recommender systems. The most important criterion in recommender systems is precision. The precision measures the accuracy of recommendations comparing to real votes. As a measure of precision evaluation, we used the Mean Absolute Error (MAE). This metric computes the mean of absolute error between predicted ratings and the real ratings that are actually assigned by users. Pn i=1 v(i) − P red(ua ,i) (6) M AE = n where v(i) − P red(ua ,i) denotes the absolute error between a real vote vi and a predicted vote P red(ua ,i) concerning an active item i and n represents the number of items in the test dataset. Since items that have high prediction values are the ones that are recommended to users, we use also the HMAE (High MAE) metric [23] to evaluate the performance of the model. The HMAE is similar to MAE but it considers only items that are predicted with a value of 4 or 5. In our experimentation, we choose the HMAE metric to measure how our system is able to recommend relevant items to active users. For both metrics, the lower the MAE and HMAE values are, the more the generated recommendations are accurate. V. R ESULTS In order to analyze the performance of BNCF, we evaluate the precision of generated predictions in terms of MAE and HMAE. The BNCF accuracy is compared to Classical CF used by standard recommender systems and to the Navigational based CF presented in [24], where only direct neighbors (that share similar navigational patterns with the active user) are integrated to generate predictions. The goal of this evaluation is to study the impact of using both neighborhoods generated by the navigational technique and the transitive associations. We note that, before the computation of predictions, in order to select reliable similarity values, we used at the same time (for all the models) two criteria to select the nearest neighbors of the active user: • The threshold (related to the similarity value) has been set to 0.2. • The minimum number of co-rated items between the active user and other users has been set to 20 [25]. Let us notice that the application of transitivity on the studied dataset leads to the enhancement of BNCF identified neighbors by about 4% comparing to the Navigational based CF.
TABLE I MAE VALUES OF COMPARED MODELS CF Models Classical CF Navigational based CF BNCF
MAE 0.763 0.789 0.782
TABLE II HMAE VALUES OF COMPARED MODELS CF Models Classical CF Navigational based CF BNCF
HMAE 0.541 0.501 0.468
A. MAE Table I presents the MAE values related to the Classical CF, the Navigational based CF and BNCF. We can first notice that accuracy only slightly decreases if we compare Classical CF (that exploits ratings to evaluate similarity values) to the Navigational based CF (that uses navigational patterns to evaluate similarity values). This confirms the idea that navigational patterns are almost as informative as rating data and may contain complementary information to ratings in order to evaluate correlations between users. Besides, the performance of BNCF is approximately similar to the Navigational based CF. The use of additional links leads to a slight improvement compared to the use of only direct links.
Contrary to Classical CF systems, no explicit preferences need to be provided by users. Preferences are inferred from users navigational activities. Besides, unlike classical usage predictive systems, BNCF is user-based and attempts to harness usage traces in order to identify behavioral similarities between users. The more two users share long common navigational patterns, the more they are correlated. Then, these behavioral similarities are employed to model relationships between users throughout a behavioral network. The objective consists in discovering new neighbors thanks to the use of social networks techniques as transitive links. Therefore, the new identified neighbors are involved in the recommendation process with the objective of improving predictions accuracy. BNCF has been evaluated both in terms of MAE and HMAE and has been compared to other CF models in order to study the impact of the navigational patterns and the transitive links on accuracy. The experimentation shows the relevance of integrating both direct and non direct neighbors identified in the behavioral network, on the accuracy of recommendations in terms of HMAE. As a future work, we intend to exploit additional techniques used in social networks and other algorithms developed in associative information retrieval as spreading activation techniques and evaluate the impact of its combination with the navigational based CF. Besides, we plan to study the combination of social networks and behavioral networks and examine its performance on recommendations precision.
B. HMAE Let us recall that only items with high prediction values are suggested by recommender systems to the active user. Thus, we are interested in evaluating the performance of studied CF models while generating high predictions. The results of this evaluation are presented in Table II. We can first notice that the Navigational based CF reaches a better HMAE comparing to the Classical CF. This means that the use of navigational patterns leads to an improvement of HMAE. Moreover, the application of transitivity in oder to discover new neighbors contributes to an important improvement of accuracy. Indeed, we observe that BNCF outperforms the Classical CF by 13% and the Navigational based CF by 7% in terms of HMAE, contrary to MAE. This improvement can be explained by the fact that when transitivity is applied, users are not connected only according to the way they have similarly rated items commonly seen, as in Classical CF. Users are joined when they share common neighbors. Moreover, this approach has the advantage to not only consider commonly rated items, but all the items users have rated. VI. C ONCLUSION In this paper, we presented a new CF approach based on a behavioral network that exploits navigational patterns and transitive links to explore associations between users.
VII. ACKNOWLEDGMENT We would like to thank Mr. Jean Philippe Blanchard and acknowledge the financial support to this project provided by the Credit Agricole Banking Group. R EFERENCES [1] M. Jamali and H. Abolhassani, “Different aspects of social network analysis,” in Proceedings of the 2006 IEEE/WIC/ACM International Conference on Web Intelligence, 2006. [2] R. Burke, “Hybrid recommender systems: Survey and experiments,” User Modeling and User-Adapted Interaction, vol. 12, no. 4, pp. 331– 370, 2002. [3] M. Claypool, P. Le, M. Waseda, and D. Brown, “Implicit interest indicators,” in Proceedings of ACM Intelligent User Interfaces Conference, 2001. [4] M. Vozalis and K. Margaritis, “On the enhancement of collaborative filtering by demographic data,” Web Intelligence and Agent Systems: An International Journal (WIAS), vol. 4, no. 2, pp. 117–138, 2006. [5] G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: A survey of the state-of-the-art,” IEEE transactions on knowledge and data engineering, vol. 17, no. 6, pp. 734–749, 2005. [6] S. Anand and B. Mobasher, “Intelligent techniques for web personalization,” Lecture Notes in Artificial Intelligence, vol. 3169, pp. 1–36, 2005. [7] P. Chan, “A non-invasive learning approach to building user profiles,” Web Usage Analysis and User Profiling, 1999. [8] M. Gery and H. Haddad, “Evaluation of web usage mining approaches for user’s next request prediction,” in Proceedings of the 5th ACM international workshop on Web information and data management. ACM Press, 2003.
[9] B. Mobasher, H. Dai, T. Luo, and M. Nakagawa, “Improving the effectiveness of collaborative filtering on anonymous web usage data,” in Proceedings of the workshop of intelligent techniques for web personalization, 2001. [10] M. Nakagawa and B. Mobasher, “A hybrid web personalization model based on site connectivity,” in WebKDD Workshop at KDD’2003, 2003. [11] M. Jalali, N. Mustapha, N. Sulaiman, and A. Mamat, “A web usage mining approach based on lcs algorithm in online predicting recommendation systems,” in Proceedings of 12th conference of information visualisation, 2008. [12] A. Banerjee and J. Ghosh, “Clickstream clustering using weighted longest common subsequences,” in Proceedings of the Web Mining Workshop at the 1st SIAM Conference on Data Mining, 2001. [13] M. Eirinaki, M. Vazirgiannis, and D. Kapogiannis, “Web path recommendations based on page ranking and markov models,” in Proceedings of the 7th annual ACM international workshop on Web information and data management. ACM Press, 2005. [14] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee, “Measurement and analysis of online social networks,” in Proceedings of the 7th ACM SIGCOMM conference on Internet measurement. ACM, 2007. [15] D. Crandall, D. Cosley, D. Huttenlocher, J. Kleinberg, and S. Suri, “Feedback effects between similarity and social influence in online communities,” in Proceeding of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2008. [16] R. Zheng, F. Provost, and A. Ghose, “Social network collaborative filtering,” IOMS: Information Systems Working Papers, vol. CeDER07-04, 2007. [17] Y. Wang, W. Dai, and Y. Yuan, “Website browsing aid: A navigation graph-based recommendation system,” Decision Support Systems, vol. 45, no. 3, pp. 387–400, 2008. [18] Z. Huang, W. Chung, T. Ong, and H. Chen, “A graph-based recommender system for digital library,” in Proceedings of the 2nd ACM/IEEECS joint conference on Digital libraries. ACM, 2002. [19] Z. Huang, H. Chen, and D. Zeng, “Applying associative retrieval techniques to alleviate the sparsity problem in collaborative filtering,” ACM Transactions on Information Systems (TOIS), vol. 22, no. 1, pp. 116 – 142, 2004. [20] I. Esslimani, A. Brun, and A. Boyer, “A collaborative filtering approach combining clustering and navigational based correlations,” in Proceedings of the 5th International Conference on Web Information Systems and Technologies (WEBIST). INSTICC Press, 2009. [21] S. Castagnos, “Mod´elisation de comportements et apprentissage stochastique non supervis´e de strat´egies d’interactions sociales au sein de syst´emes temps r´eel de recherche et d’acc´es a` l’information,” Ph.D. dissertation, Nancy 2 University, France, 2008. [22] J. Herlocker, J. Konstan, A. Borchers, and J. Riedl, “An algorithmic framework for performing collaborative filtering,” in Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval, 1999. [23] L. Baltrunas and F. Ricci, “Dynamic item weighting and selection for collaborative filtering,” in Web mining 2.0 Workshop, ECML-PKDD 2007. Springer-Verlag, 2007. [24] I. Esslimani, A. Brun, and A. Boyer, “Enhancing collaborative filtering by frequent usage patterns,” in Proceedings of the First International Conference on the Applications of Digital Information and Web Technologies. ICADIWT, 2008. [25] P. Viappiani, B. Faltings, and P. Pu, “Preference-based search using example-critiquing with suggestions,” Journal of artificial intelligence Research, vol. 27, pp. 465–503, 2006.