Recommender systems based on social networks

Highlights

• We propose a regularization approach that incorporates social network information.

• Biclustering algorithm is used to calculate the number of realistic friends.

• The friendships are used to calculate the similarities between users.

• The correlation between user and item is taken into consideration.

• We examine the impacts of parameters to the experimental results.

Abstract

The traditional recommender systems, especially the collaborative filtering recommender systems, have been studied by many researchers in the past decade. However, they ignore the social relationships among users. In fact, these relationships can improve the accuracy of recommendation. In recent years, the study of social-based recommender systems has become an active research topic. In this paper, we propose a social regularization approach that incorporates social network information to benefit recommender systems. Both users’ friendships and rating records (tags) are employed to predict the missing values (tags) in the user-item matrix. Especially, we use a biclustering algorithm to identify the most suitable group of friends for generating different final recommendations. Empirical analyses on real datasets show that the proposed approach achieves superior performance to existing approaches.

Introduction

Recommender systems have attracted much attention in the past decade. A recommender system is a software tool that supports users in identifying the most interesting items. There has been much work done on developing new approaches to recommender systems (Adomavicius and Tuzhilin, 2005, Brusilovsky and David, 2013). The research topic is still popular because of the abundance of practical applications that help users to deal with information overload and their great commercial value. Examples of the applications include recommending books, movies and some other commercial systems.

With the development of Web 2.0, the study of social-based recommender systems started. The traditional ones (Adomavicius and Tuzhilin, 2005, Brusilovsky and David, 2013) always ignore social interactions among users which can improve recommender systems. The fact is, when we are confused by multiple choices, we may turn to our related friends for the best recommendations, since they are those who we can reach for immediate advice. Hence, in order to provide more accuracy and personalized recommendation results, the social network information should be incorporated. Based on the above viewpoints, a few trust-based recommender systems (Jamali and Ester, 2010, Massa and Avesani, 2004, Ozsoy and Polat, 2013, Massa and Avesani, 2009, Bedi et al., 2007, Nazemian et al., 2012, Ma et al., 2008) which move an important progress forward have been proposed. The methods utilize the unilateral trust information to further improve traditional recommender systems. However, these methods have several inherent limitations and weaknesses that need to be solved. The noticeable weakness is the unilateral “trust relationship” problem. It is different from the concept “social relationship” which refers to the cooperative and mutual relationship between users. In addition, the other weaknesses are the impracticable hypothesis and the weak generalization ability. Obviously, the trust-based recommender systems may no longer be suitable. Therefore, the study of real social-based recommender systems appears on the screen. Additionally, the integration of social networks can theoretically improve the performance of traditional recommender systems. First, in terms of the prediction accuracy, the friendships among users improve the understanding of user ratings. Therefore, we can interpret user preferences more precisely. Second, as a matter of fact, the friendship between two users already indicates that they have things in common. Thus, the cold-start problem can be alleviated (Jamali and Ester, 2010, Massa and Avesani, 2004).

In order to solve the problems mentioned above, in our research, we focus on the social-based recommender system and, similar to Ma et al. (2011), propose an approach named RSboSN (Recommender Systems based on Social Networks) that integrates social network graph and the user-item matrix to improve the prediction accuracy of the traditional recommender systems. In the process of recommendation, we mainly use the friendships among users and the tags labeled by the users to recommend. The user-item-tag can be considered as a two-dimensional matrix. We cluster the similar users to calculate the similarity between users and the correlation between a user and an item. The purpose in clustering is to identify the most suitable friends for realistic recommendation tasks. Based on the approach in Ma et al. (2011), the above two detailed aspects of social network information are employed in designing social regularization terms. We also take the situation into consideration that different friends may have dissimilar or even opposite tastes. Even if the friends of the same group focus on the same item, they may have different favorite degree. We have conducted experiments on real dataset to evaluate the performance of our approach on the prediction accuracy. The experiments show significant improvement over traditional and state-of-art social-based recommender systems in those aspects.

The remainder of the paper is organized as follows. Section 2 presents the overview of related work. Section 3 defines the problem and presents the details of the proposed approach. Section 4 presents the experiments results. Finally, we draw the conclusion in Section 5.