Factor-reduced matrix factorization with a new initialization method

Abstract. Recommender systems have become an indispensable component of digital life. At the outset of the second quarter of the twenty-first century, one of the most effective approaches for their implementation is matrix factorization (MF), a learning algorithm whose input is a matrix generated by users through rating items with which they are familiar. MF is a latent factor method: although the algorithm does not explicitly observe user and item attributes during training, it can nevertheless predict the extent to which an unknown item is expected to satisfy a given customer’s preferences. In real-world applications, additional information is often available, from which certain item properties can be inferred. Incorporating item attributes into MF is commonly achieved by setting the initial values of the item feature matrix accordingly. In this paper, we present a method that leverages item attributes to generate suitable initialization values for the user feature matrix, thereby improving the predictive accuracy and convergence speed of the learning algorithm.

Key words and phrases. Rating-based recommender system, hybrid matrix factorization, autoencoder.