2018 G. Satya Keerthi PhD Thesis Awarded
A DISTRIBUTED COMMUNITY BASED SOCIAL RECOMMENDER APPROACH USING TRUSTED NEIGHBOURHOOD
Abstract: Recommender Systems (RS) apply knowledge discovery techniques on the fly for making personalized recommendations related to information, products or services. These systems, especially the k-nearest neighbour collaborative filtering based ones, are used widely on the web. The tremendous growth in the amount of available information and the number of visitors to websites in recent years pose some key challenges for recommender systems. The challenges are producing high quality recommendations, performing many recommendations per second for millions of users and items while achieving high coverage in the face of data sparsity. In traditional collaborative filtering systems the amount of processing increases with the number of participants in the system. New recommender system technologies are needed that can quickly produce high quality recommendations, even for very large-scale problems. To address these issues we have explored community-based content and collaborative filtering techniques in this thesis work. Collaborative techniques first analyze the co-occurrences between items to identify relationships. These relationships are then used to form communities and compute recommendations for users based on the active users’ target item. We look into different community detection techniques and explored a new distributed community approach in order to reveal the topological relations in the network. However, the user-item ratings matrix, which is used as input to the recommendation algorithm, is often highly sparse, leading to unreliable predictions. Recent studies demonstrated that information from social networks such as trust statements can be employed to improve accuracy of the recommendations. However, there are explicit trust relationships between most of users in many e-commerce applications. In this thesis, a method to identify implicit trust statements by applying a reliability measure is proposed. Finally, the results are experimentally evaluated and compared with the traditional recommender approaches. The experimental results showed that community-based algorithms provide dramatically better performance than User-Based Collaborative Filtering (UBCF) algorithms and Item-Based Collaborative Filtering (IBCF) algorithms, while providing better quality than the available user-based algorithms.