Collaborative Filtering Item Recommendation Methods based on Matrix Factorization and Clustering Approaches

Abstract

This paper studies various collaborative filtering item recommendation methods based on matrix factorization and clustering approaches. We develop six methods that are modified based on three matrix factorization approaches, i.e., Non-Negative Matrix Factorization (NMF), Singular Value Decomposition (SVD), and Principal Component Analysis (PCA); and two clustering techniques, i.e., K-Means and Fuzzy C-Means. The framework of method development consists of three main phases: the matrix factorization for generating the latent factors; the users clustering for grouping similar users; and the user-based collaborative filtering for generating the item recommendation. The recommendation performances are evaluated in terms of Fl-Score and Normalized Discounted Cumulative Gain (NDCG) metrics, at various top-N. Using the Movie Lens rating dataset, experiment results show that the combination of PCA and K-Means outperforms the other five methods, where the global averages of outperformance are 82.05% in terms of Fl-Score and 52.10% in terms of NDCG. It is also worthwhile to note that the combination of K-Means clustering with any matrix factorization techniques is superior compared to that of the Fuzzy C-Means. Additionally, the advantage of implementing the matrix factorization approach depends on which clustering technique used in the method.