Інтеграція контентних ознак при ініціалізації латентних факторів в системах персоналізованих рекомендацій

Abstract

The study examines an approach to improving matrix factorization by integrating content-based features during the initialization of latent factors. The limitations of the standard SVD method, particularly the “cold start” problem, are analyzed, and an integration of content-based and context-based models is proposed, combining factorization with vectorization of content characteristics. It is shown that using the TF-IDF method for initial parameterization enhances prediction accuracy and accelerates the convergence of the training procedure.