Preliminary data-based matrix factorization approach for recommendation

Abstract

Existing collaborative filtering algorithms suffer from the problem of data sparsity. Imputation-based methods are promising algorithms, which alleviate data sparsity without using side information, to solve this problem. However, existing imputation recommendation methods based on matrix factorization only separately factorize the rating matrix and the imputed data matrix, which limits the power of imputed data. In this paper, we propose an efficient method, which can make full use of the imputed data, to alleviate data sparsity. Firstly, our method, called Preliminary Data-based Matrix Factorization (PDMF), generates preliminary prediction data based on neighborhood-based methods. Secondly, PDMF consists of two models of learning the user and item preferences. One firstly makes the original preferences get close to preliminary preferences, and then creates the concatenated preferences. The other one creates the concatenated preferences firstly, and then makes the original, preliminary and concatenated preferences get close to each other. To the best of our knowledge, our method is the first to constrain the learning procedure in matrix factorization by using imputed data. We test our method on the MovieLens100k, MovieLens1M, Netflix, Filmtrust and Jester datasets. Experiment results show that the PDMF outperforms the state-of-the-art methods in recommendation accuracy.