In this paper, the rating scale model is extended from one-dimension to multi-dimension, and then, a novel collaborative filtering algorithm is proposed. In this algorithm, user’s interest is multi-dimensional, and item’s quality that satisfies user’s interest is multi-dimensional too. The rating of a user for an item is a weighted summation of all the latent ratings of the user for the item in all dimensions, and the weights at different interest dimensions are user-specific. The latent rating of user u for item i in one dimension is of a multinomial distribution which is determined by the user’s interest value in this dimension, the item’s quality value in this dimension, and the user’s rating criteria. The parameters are estimated by minimizing the loss function using stochastic gradient descent method. Experimental results on benchmark datasets show that the algorithm has better performance than the compared algorithms.
展开▼
