Social recommendation based on users’ attention and preference
Introduction
Social recommendation has attracted considerable research attention in recent years as it effectively addresses data sparsity and cold start issues in traditional recommendation algorithms. One major assumption in social recommendation is that users’ behaviors in a social network are heavily affected by their social relations. Thus it is possible to infer users’ preference via their social relations for more precise recommendation. However, existing methods in social recommendation mainly generate recommendations based on users’ preference and overlook the importance of users’ attention. Besides, we observe that the influence of social relations dwells more on users’ attention than on their rating values in four real-world datasets.
Attention is the behavioral and cognitive process of selectively concentrating on small fraction of information, while ignoring other perceivable information [1], [2], [3]. Therefore our attention is selective and limited by its nature [4], [5], [6]. Users tend to capture information that supports their biased pre-existing views from their prior experience and social network [7], [8]. Thus it is not surprising when presented a list of recommendations, we will focus more on the items that draw our attention and skip some other items even if they suit our tastes. This psycho-social effect motivates us to model users’ attention in social recommendation for better recommendation accuracy.
To better understand the characteristics of users’ attention in social recommendation, we conduct empirical analyses based on four large real-world datasets in this paper. Specifically, in recommendation, we explore users’ attention based on their attention behavior, i.e. which items the user have spent mental effort on consuming and rating. An important conclusion can be drawn: Users’ attention behavior is heavily affected by their trust relations, while their rating values stay relatively indifferent to them. Thus, on the one hand, social information can be used more effectively to infer users’ attention than users’ preference. On the other hand, each of us belongs to some content-shared communities and our attention will inevitably be influenced by our social relations [2], [9], which makes that our attention is not coincident with our preference. Thus, just considering one aspect for recommendation may not produce effective recommendation results. Specifically, when presented a list of recommendations, we usually focus more on the items that draw our attention, and are unwilling to spend effort to learn the information about the rest. This means that the recommended items just suiting our tastes may be ignored because of our limited and selective attention [4], [5]. Meanwhile, the items which just draw our attention may not be accepted, since when we make the decisions on the consumption, we usually observe ourselves and judge whether we will like them. Therefore, to improve the recommendation accuracy, in modern recommender systems, how items draw users’ attention and how items suit users’ tastes should be both taken into consideration.
Based on the above analyses, we propose a new social recommendation model HTPF that explicitly considers both users’ attention and preference for better recommendation accuracy. Besides, different from existing methods that model social influence on users’ rating values, HTPF differs from these methods in that it uses social information to better infer users’ attention, which is more suspectable to social network. Considering Poisson factorization [10] is particularly efficient on sparse data and usually achieves better performance than traditional methods (such as PMF [11], similar-based methods [12], random walk [13]), HTPF employs an integrated latent factor model based on Poisson factorization [10]: HTPF fuses the Poisson factorization of attention behavior information and trust information by sharing common latent vectors of users’ attention, since users who have similar social relations tend to pay attention to similar items. At the same time, HTPF fuses the Poisson factorization of attention behavior information and rating information by sharing common latent vectors of items, because of constant item attributes. A scalable variational inference algorithm for our HTPF model has been developed to predict how the specific items draw users’ attention and how the items suit users’ tastes. Targeting at high recommendation accuracy, HTPF generates recommendations based on the weighted combination of these two aspects. We also conduct extensive experiments and demonstrate that by combining users’ attention and preference our method outperforms state-of-the-art social recommendation methods.
It is worthwhile to highlight the following contributions of our work:
- •
We introduce user’s attention in the social recommendation and uncover that the influence of trust relations dwells more on users’ attention than on their preference.
- •
We propose a novel probabilistic model HTPF that explicitly models both user’s attention and preference for social recommendation. Besides, We develop an efficient inference method for HTPF.
- •
Our experiments on real-world datasets clearly demonstrate the superiority of our method over existing social recommendation methods.
Before proceeding further, we now formalize the problem definition. Suppose we have a recommender system with users set U (including n users) and items set I (including m items). Rating information is represented as n × m matrix R whose entries (Rui) denote the rating value of item i given by user u. Attention behavior information indicates observed attention behaviors that users are willing to spend mental effort on consuming and rating some specific items [1], [14], represented as n × m matrix Ω. For each entry in Ω, denotes that user u have consumed and rated item i, and means not. Trust information indicates the trust relations between users, represented as n × n matrix G. For each entry in G, denotes user u trusts user v, denotes not (Note that we can treat undirected friend relations as bi-directed trust relations in this paper). The task of social recommendation is to give each user an item list that will be accepted with pleasure based on these information.
The rest of this paper is organized as follows. we briefly review related works in Section 2. The social empirical analyses are conducted in Section 3. In Section 4, we introduce the details of our proposed model HTPF. The experiment results and discussions are presented in Section 5. Finally, we conclude the paper and present some directions for future work in section 6.
Section snippets
Related works
With the exponential growth of information generated on consumer review websites and e-commerce websites, recommender systems are drawing more attention from both academia and industry [15], [16], [17], [18], [19], [20], [21]. For the system with explicit feedback (numerical ratings), substantial works have been done about collaborative filtering (CF) model for its accuracy and scalability during the past two decades [22], [23], [24], [25], [26]. To address the limitation of the traditional CF
Analyses of social influence on datasets
In this section, we conduct empirical analyses on four large real-world datasets: Epinions1, Ciao2, Flixster3and Douban4. The four datasets contain several kinds of information. The rating information denotes the rating values that the user gave to some items. The rating values in Epinions, Ciao and Douban are
Model
In this section, we present our proposed model HTPF and its inference algorithm.
Experiments and results
In this section, we evaluate our algorithm on real-world datasets. We start with the description of four datasets.
Conclusions
In this paper, we propose a new probabilistic model HTPF that explicitly considers both users’ attention and preference in social recommendation. Many psycho-social literatures suggest the importance of users’ attention in recommendation and our observations in Section 3 show that the influence of trust relations dwells more on users’ attention than on their preference. Thus, we propose the model HTPF with a generative process where we use social network as complemental information to deduce
Acknowledgemnt
This work is supported by the National Natural Science Foundation of China (Grant no: U1866602) and by a Discovery Grant from the National Science and Engineering Research Council of Canada. It is also partially supported by ByteDance.
Jiawei Chen received his B.S. at University of electronic science and technology of China in 2014. he is currently a Ph.D. candidate in the college of computer science at Zhejiang University. His main research topics are recommender systems, graphical model and social networks.
References (46)
- et al.
Selective exposure
Adv. Exp. Soc. Psychol.
(1965) - et al.
Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions
IEEE Trans. Knowl. Data Eng.
(2005) Cognitive Psychology and its Implications
(1990)Attention and Effort
(1973)- et al.
Load theory of selective attention and cognitive control.
J. Exp. Psychol.: Gen.
(2004) - et al.
To see or not to see: the need for attention to perceive changes in scenes
Psychol. Sci.
(1997) - et al.
LA-CTR: a limited attention collaborative topic regression for social media
Proceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence
(2013) - et al.
Learning to selectively attend
Proceedings of the 32nd Annual Conference of the Cognitive Science Society
(2010) - et al.
Learning and selective attention
Nat. Neurosci.
(2000) - et al.
Uncovering the overlapping community structure of complex networks in nature and society
(2005)
Scalable recommendation with hierarchical poisson factorization
Proceedings of the Thirty-First Conference on Uncertainty in Artificial Intelligence, UAI 2015, July 12-16, 2015, Amsterdam, The Netherlands
Probabilistic matrix factorization
Proceedings of the Advances in Neural Information Processing Systems
Item-based collaborative filtering recommendation algorithms
Proceedings of the 10th international Conference on World Wide Web
Trustwalker: a random walk model for combining trust-based and item-based recommendation
Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Gated probabilistic matrix factorization: learning users’ attention from missing values
Proceedings of the International Joint Conference on Artificial Intelligence
Recommender systems: introduction and challenges
Recommender Systems Handbook
Deep learning based recommender system: a survey and new perspectives
CoRR
A brand-level ranking system with the customized attention-GRU model
Proceedings of the 27th International Joint Conference on Artificial Intelligence
What to do next: modeling user behaviors by time-LSTM
Proceedings of the 26th International Joint Conference on Artificial Intelligence
Which to view: Personalized prioritization for broadcast emails
Proceedings of the 25th International Conference on World Wide Web
The million domain challenge: Broadcast email prioritization by cross-domain recommendation
Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Heterogeneous hypergraph embedding for document recommendation
Neurocomputing
Factorization meets the neighborhood: a multifaceted collaborative filtering model
Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Cited by (34)
Integrating interactions between target users and opinion leaders for better recommendations: An opinion dynamics approach
2023, Computer CommunicationsCitation Excerpt :Users often ask their friends on social media for information or opinions about products and services [4,5], which means that users’ social network relationships can help them to filter information. Therefore, scholars have attempted to integrate users’ social relationship information into the traditional recommendation framework in order to better recommend items to the target user who want to find items [9–14]. Various social recommendation algorithms based on the matrix factorization model have been proposed [15–17], including Soreg [18], Sorec [19], SocialMF [20], RSTE [21].
Adaptive social recommendation combined with the multi-domain influence
2023, Information SystemsCitation Excerpt :The model integrates the user–item interaction graph and social graph to obtain the user’s latent vector, and mines the inherent features of the user–item interaction behavior to make recommendations. In HTPF [26], a probabilistic model is used to consider both users’ attention and preferences in social recommendation, and it is concluded that it is more effective to use social relations to infer user attention. Li et al. [27] consider the heterogeneity of online social networks and users’ trust in friends vary from item to item, and proposes a social recommendation method based on co-attention neural network.
URPI-GRU: An approach of next POI recommendation based on user relationship and preference information
2022, Knowledge-Based SystemsTowards comprehensive profile aggregation methods for group recommendation based on the latent factor model
2021, Expert Systems with ApplicationsCitation Excerpt :Many previous studies in the group recommendation have shown that profile aggregation usually produces slightly better performance than recommendation aggregation (Baltrunas, Makcinskas, & Ricci, 2010; Berkovsky & Freyne, 2010a; Boratto & Carta, 2015; Lin, Chiu, & Chen, 2011; Nguyen & Ricci, 2018; Pessemier, Dooms, & Martens, 2013; Popescu & Pu, 2012; Quan & Cho, 2014; Van Deventer, De Wit, Vanattenhoven, & Gualbahar, 2013; Wang et al., 2016). For profile aggregation, the aggregation functions that have the contribution of more and more group members are more reliable (Chen et al., 2019; Felfernig et al., 2018; Guo et al., 2020; Masthoff, 2011; Wang, Chen, et al., 2020; Xiao et al., 2020; Yalcin & Bilge, 2021). From this point of view, a weighted average function is particularly suitable for aggregating profiles of group members (Berkovsky et al., 2010b; Felfernig et al., 2018; Nguyen & Ricci, 2018; Yalcin & Bilge, 2021).
An appraisal mechanism for a social marketplace
2021, Information and ManagementJoint representation learning with ratings and reviews for recommendation
2021, NeurocomputingCitation Excerpt :However, the rating data is usually highly sparse, leading to that the relevance between users/items is hard to be discovered. To address this problem, various kinds of auxiliary information have been introduced into matrix factorization, such as tags [4–6,40,41], social relations [7,8,42], reviews [9–11,43], and visual features [12,13,35–37]. In this auxiliary information, the reviews generated from the user behaviors can reflect user preference directly, and it is easy to provide reasonable explanations for recommendation results.
Jiawei Chen received his B.S. at University of electronic science and technology of China in 2014. he is currently a Ph.D. candidate in the college of computer science at Zhejiang University. His main research topics are recommender systems, graphical model and social networks.
Can Wang received the Ph.D. degree and M.S. degree in computer science and B.S. degree in economics from Zhejiang University, in 2009, 2003 and 1995, respectively. His research interests include data mining, machine learning and information retrieval.
Qihao Shi received his B.S. at Nanjing Normal University of China in 2014. He is currently a Ph.D. candidate in the college of computer science at Zhejiang University. His main research topics are social and information networks, algorithmic game theory and Internet economics.
Yan Feng received the Ph.D. degree in computer application from Zhejiang University in 2004. She is currently an associate professor in the College of Computer Science at Zhejiang University, China. Her research interests include database, data mining etc.
Chun Chen is a professor in the College of Computer Science, Zhejiang University. His research interests include data mining, computer vision, computer graphics and embedded technology.