Latent Factor Models Fusing User & Item Attributes

Abstract

Data sparsity, cold-start, and suboptimal recommendation for local users or items have been recognized as the most crucial three challenges in the latent factor model (LFM) for recommender systems. This paper proposes an approach that integrates the User-Item attributes into the classical LFM named UILFM focusing on above challenges. First, for the problem of data sparsity and cold-start, we develop an online learning algorithm to update the weights of user or item attribute for identifying the importance of different attributes. By aggregating the users and items based on their similar attributes, we obtain the local neighbor group which makes it possible for recom- mender to estimate some missing ratings based on adjacent user's ratings towards items and adjacent item's ratings. By introducing the convex mixed-parameters, we combine the estimate ratings with the classical LFM to predict the missing entries of the high-dimensional and sparse (HiDS) matrix for further closing the true ratings and reducing matrix sparsity. Second, for the suboptimal recommendation problem, we propose a new matrix filling (for missing ratings) method based on positive and negative samples, in which when the sparsity of the HiDS matrix is reduced to a threshold, the classical LFM will dominate the filling procedure, instead, the prediction based on neighbors' ratings remains a domination role. This method elegantly solves the suboptimal recommendation problem that the ratings of partial users are extremely sparse and the number of ratings per user are unbalanced. The proposed algorithm is tested by the MovieLens dataset, the results show that it promotes the recommendation accuracy compared with the classical LFM algorithm and the dimensionality reduction approaches as well as the collaborative filtering (CF) algorithms.