Effective Approaches to Collaborative Filtering and Its Application

Abstract

In e-commerce era, recommender systems have been widely used to help users in finding the items in which they might be interested. Collaborative filtering (CF), one of the most popular and effective techniques of recommender systems has been widely studied for the last decade. In this dissertation, we study three interesting problems of CF and provide effective solutions to them. In the first part of the dissertation, we study the problem of making recommendations to a group of users. The accuracy of existing approaches is significantly affected by the size and cohesiveness of a user group. We propose a novel approach that makes effective recommendations to a group of users regardless of its size or cohesiveness. We first model the relationships between a set of users and a set of items as a bipartite graph from the rating information. On this graph, we employ the belief propagation algorithm to determine probabilistically the target user group’s preferences on items. We also propose a new group type that reflects real-life groups effectively and helps in the better evaluation. Extensive experiments on synthetic and real-world datasets show that the proposed approach is more accurate up to 20% than the existing ones. In the second part of the dissertation, we study the well-known cold-start user problem of CF. Several approaches exist in the literature that aim to solve the cold-start user problem of CF. Most of them exploit only trust or both trust and distrust relationships. We point out four problems of existing approaches: sparsity of explicit signed trust (i.e., trust and distrust) relationships, the existence of dissimilar/similar preferences between user pairs having trust/distrust relationships, high computational cost, and the problem due to employing the transitivity of distrust relationships. We propose a novel approach effectively exploits both explicit signed trust relationships and the similarity scores between users’ ratings to solve the aforementioned problems. The proposed approach infers implicit signed trust relationships between users, confirms the degree of similarity/dissimilarity of user pair having trust/distrust relationships, selects top-k signed relationships of a user, and employs the intransitivity of distrust relationships. Extensive experiments on a real-world dataset show that the proposed approach outperforms existing ones up to 10% in terms of accuracy. In the third part of the dissertation, we study a novel problem of viral marketing through a CF-based recommender system. Existing studies show that both optimal and approximation solutions to the problem of viral marketing through a CF-based recommender are NP-hard within any reasonable factor. We solve the aforementioned problem by proposing, (i) a novel diffusion model that effectively supervises the influence spread in a CF-based recommender system; (ii) a novel problem that aims at viral marketing through CF-based recommender system; and (iii) an approximate solution to the proposed problem that is provable within 63% of the optimum. Extensive experiments on three real-world datasets show that seed users selected by the proposed algorithm consistently outperform the seed sets selected by all other competitors in terms of promoting the target item to other users.