Abstract: Matrix factorization has become a common way to predict user ratings of items. Traditional matrix factorization algorithms do not take account of the differences between users. To address this problem, a difference value(D-value) matrix factorization model is proposed. First, for each user, the difference between his/her rating score and the average rating score from users with similar social attributes is calculated, which finally results in a matrix called D-value matrix. Then the D-value matrix is factorized to calculate the predicted ratings. Experimental results on the Movielens 1M dataset show that the proposed method significantly outperforms Bayesian probabilistic matrix factorization, matrix factorization and the latent factor model fused with user attributes in terms of prediction accuracy.


Key words: recommendation algorithms, matrix factorization, D-value matrix factorization