A novel 2D-Graph clustering method based on trust and similarity measures to enhance accuracy and coverage in recommender systems

Various clustering approaches have been widely adopted to improve the accuracy and scalability of collaborative filtering-based recommender systems as the major objectives. Recent research has experimentally disclosed the realization of such objectives at the expense of decreasing the recommendations coverage. In this paper, we prove that the rate of coverage shows a monotonic decreasing trend as the number of clusters increases. This proof establishes a foundation for the proposal of a 2D-Graph clustering method for partitioning a novel 2D-Graph in compliance with the minimum-acceptable-rate-of-coverage criterion. To reduce the negative impact of clustering on the coverage in the proposed method, in addition to the information of trust, Pearson similarity is incorporated to construct double-weighted connections between users in the 2D-Graph. In each cluster, to address the sparsity problem, the weights corresponding to a convex combination of Pearson, trust and Jaccard similarities are determined so that the precision is optimized. The experimental results on two real-world datasets illustrate that the proposed method outperforms the state-of-the-art methods for Cold Start users in terms of accuracy and coverage.