רועי אנואר

Abstract :

Matrix factorization is a popular method for modeling user and item behaviors in recommender systems. Matrix factorization is used to capture the item and user biases when building a recommendation model, thus representing the User-Item-Rating space in a compact form. One of the most known methods used for matrix factorization is SVD (Singular Value Decomposition). With SVD, the usage of matrix factorization results in a representation of the original user-item-rating matrix as a combination of latent (implicit) dimensions, which are perceived as the user interests and the item attributes. Understanding these dimensions and their translation to true user preferences is a difficult task. Our research focuses on non-latent (explicit) factorization of the user-item-rating matrix. We show that, given existing item attributes, we may build an explicit model describing the user preferences. Using this model we demonstrate how to predict ratings using explicit attributes from the data. To select the best attribute for our model, we utilize the realization based entropy approach and define two new measures: (1) the space entropy and (2) the consent entropy. Using a combination of these entropies, we may quantify the contribution of each non-latent attribute to our prediction model. Using the Movielens datasets, we compare our results to the known "SVD" model, and show how our explicit model yields similar results.