Trend Analysis with Interpretability and Cold-Start Problems for Recommender Systems

Matrix factorization is a common method in recommender systems. However, distinguishing between continuously and temporarily popular items is challenging because basic matrix factorization relies on the accumulated user rating records for an item. Moreover, recent trends emphasize that recommender systems should be both accurate and interpretable, necessitating clear reasons behind each recommendation. In this paper, we propose temporal positive collective matrix factorization (TPCMF), which improves the interpretability and temporality of collective matrix factorization. We make the factor matrix obtained by collective matrix factorization non-negative to increase the interpretability. In addition, we take into account the time variations of the factor matrices and make time-series predictions, which enables temporal recommendations. Moreover, in experiments using real-world datasets, we determined that factor interpretation under TPCMF provides substantial insights into the interpretability and temporality of recommendations with accuracy surpassing existing methods. In addition to these features, we also propose a preprocessing method to address the cold-start problem, which is a common issue in matrix factorization.