Trust-Based Item Recommendation Methods In E-Commerce

In recent years, with the rapid development of internet technology, information has in-creased at an unprecedented rate and the information overload has become increasingly severe for online users. In order to solve the information overload problem, researchers and engineers have presented lots of techniques where recommender system is one of the most popular and effective tools. Nowadays, recommender system has been applied in various domains where E-commerce is one of the most important applications. In E-commerce, the increasing num-ber of users and items makes it more difficult to find items of interest. Recommender system for E-commerce can analyse behavior pattern from users'historical data and then predict users' preferences so that it is promising to solve information overload problem in E-commerce do-main. On one hand, users can receive personalized service and experiences. On the other hand, the websites in E-commerce can increase product sales and profits.However, recommender system is suffering from some challenges, such as data sparsity, cold start, and diverse recommendation, which have limited the development of E-commerce. Therefore, in order to solve those problems, in this thesis, we mainly focus on the scenario where existing methods cannot be applied and the characteristics hidden in trust relations as well as diversity metric which are ignored in existing methods, take the limitation in E-commerce (i.e., diverse product categories) into consideration, and study trust relations-based recommendation:First, we propose an items' relations mining-based cross-domain recommendation method for solving the problem of data sparsity and cold start existed in single-domain recommendation. In the cross-domain scenario where there are no common users and common items, we utilize the information on trust and ratings from source and target domains to generate representations for all items and then build the association between two domains. Based on their representations, we compute the similarities between cross-domain items for predicting unknown ratings of users from source domain on items from target domain by transferring known ratings from source domain. Relevant experimental results on real-world datasets show that, our proposed method can effectively build the associations between different domains and solve data sparsity as well as cold start.Second, we propose a matrix factorization-based recommendation method using domain-specific trust relations for solving the problem of data sparsity which exists in recommendation methods based on the use of trust relations without explicit domain. We focus on the fact that interest domains are diverse and users' trust behaviors across different domains are different. We divide interest domains of users and then construct the domain-specific trust networks based on direct and indirect trust relations. We compute the similarities between users on trust networks and construct a matrix factorization model based on the information on trust and ratings. Relevant experimental results on real-world datasets show that, our proposed method can utilize domain-specific trust behavior to improve recommendation quality and solve the data sparsity.Third, we propose a collaborative filtering based on implicit relation mining between users with the same role for solving the problem of data sparsity which exists in recommendation methods based on explicit and implicit relations between users. We analyze the co-occurrences between trusters as well as trustees and then construct the associations between users with the same role. Based on their associations, each user is represented by the role-same users to him for computing their implicit similarities. We fuse the implicit similarities into collaborative filtering for predicting accurate unknown ratings. Relevant experimental results on real-world datasets show that, our proposed method can utilize the mined implicit relations between users with the same role to improve recommendation quality and solve the data sparsity.Fourth, we propose a trust relations and re-ranking strategy-based item recommendation method for improving diversity and then satisfying E-commerce websites'requirement. Based on the associations between the creation time of trust relations and diverse recommendations, we weight trust relations created in different time points to improve diversity. Based on the as-sociations between items' popularity and diverse recommendations, we present two re-ranking strategies to represent high-popularity items in recommendation list with low-popularity items for improving diversity. Relevant experimental results on real-world datasets show that our pro-posed method can effectively utilize the creation time of trust relations and the items' popularity to improve recommendation diversity and solve the problem of diverse recommendation.