Comparison of Matrix Factorization Methods for Item-Based Recommendations

Modern recommender systems increasingly go beyond classical personalization tasks, addressing more complex scenarios of interactions between items. One such challenge is generating complementary recommendations, where standard user-centric architectures often lack sufficient flexibility. This study compares two matrix factorization-based approaches to solving this problem: a classical model trained on the user–item matrix with additional constraints derived from cooccurrence statistics, and a direct factorization of an item–item matrix constructed using a temporal coaction rule. This paper analyzes ways to overcome the limitations of traditional methods and outlines the potential of new strategies across various data types and business applications.