Traditional collaborative filtering recommendation methods calculate similarity between users to find the most similar neighbors for a particular user and take into account their opinions to predict item ratings. Though these methods have some advantages, however, they encounter difficulties in dealing with the problems of cold start users and data sparsity. To overcome these difficulties, researchers have proposed to consider social context information in the process of determining similar neighbors. In this chapter, we present a data analytics approach that combines user preference and social trust for making better collaborative recommendation. The proposed approach regards the collaborative recommendation as a classification task. It includes a data analysis procedure to explore the target dataset in terms of user similarity and trust relationship, and a data classification procedure to extract data features and build up a model accordingly. A series of experiments are conducted for performance evaluation. The results show that this approach can be used to enhance the recommendation performance in an adaptive way for different datasets without an iterative parameter-tuning process.
TopIntroduction
Recommender systems have been advocated in different service domains for years (Adomavicius & Tuzhilin, 2005; Bobadilla, et al., 2013; Lu, et al., 2015; Ricci, et al., 2011). Traditional recommender systems address two entities for application services: the users and the items. Initially, the systems collect some ratings specified by users. Based on these rating records, these systems estimate the rating function R: Users × Items → Ratings. Once the utility function R is constructed, the system can predict the rating of unfamiliar items and recommend the highest-rating ones to users. Various methods have been developed to find effective solutions that require less computational effort. These methods range from content-based user modeling to group-based collaboration (that is, collaborative filtering or CF). Generally speaking, the group-based approach is more efficient and effective than content-based user modeling (Bobadilla, et al., 2013; Koren & Bell, 2011).
Many collaborative recommender systems have tried to predict the rating of an item for a particular user based on how other users previously rated the same item. Algorithms for collaborative recommendations can be grouped into two classes in general: memory-based and model-based methods. Memory-based algorithms (also called neighborhood or k-nearest neighbor methods) are heuristics that make rating predictions based on the entire collection of items previously rated by the users. The neighborhood methods are popular because they are intuitive and relatively simple to implement. Moreover, these neighborhood methods offer useful and important properties: explicit explanation of the recommendations and easy inclusion of new ratings. Because our major goal in this work is to investigate how to incorporate different types of information and to analyze their effects on recommendation, we thus choose this easy-to-implement approach in our experiments.
Standard CF methods calculate user similarity to find neighbors and their opinions to make decisions. Though similarity-based neighborhood methods have some advantages, there also occur several issues to consider. For example, if the interactions among the neighbors are not considered, and some users have rated only a very small number of items (i.e., the cold-start problem), such information is not enough for making helpful recommendation. Also this method has a problem of sparsity: in real-world applications, most items are not widely rated by users. The latter problem may decline the recommendation performance (Aha, 2008; Guo, et al., 2014). To overcome these problems, researchers have suggested the inclusion of contextual information for building more accurate recommender systems, for example (Chen, et. al., 2012; Chen, et al., 2014; Lee and Lee, 2014; Ma, et al., 2011). One kind of contextual information, the social context information collectable from online communities (i.e., social networks), is useful and important for recognizing the users’ situation that can influence his decision (Qian, et al., 2014; Samah, et al., 2012; Yang, et. al. 2013). Social network theory suggests that the positions of users in a web of relationships influence their access to resources, friends, and information. Social influence can be used to create intention for people to consume a product so that everyone’s social relationship is one of the key factors for predicting the potential customer intention. Researchers have indicated that the relationships among friends and friends of friends within a social network are crucial when referencing trustworthy and reliable information (Golbeck & Hendler, 2006; Jamali & Ester, 2009; Ray & Mahantirt, 2010).