Abstract: Recommender system is a competitive solution to solve information overload problem.And collaborative filtering is an effective method for the recommender systems.Matrix factorization is widely used for collaborative filtering.However,the existing matrix factorization techniques are affected by the rating noise,and their robustness is not up to people’s expectations.We attributed the negative effect of the rating noise to the universal applied hypothesis that the rating data is subject to the Gauss distribution.In order to solve this problem,we proposed a collaborative filtering algorithm based on Gaussian mixture model.We assumed the rating data obeying the Gauss mixture distribution,and then applied the Bayesian probability matrix factorization model for recommendation.Besides,a semi-supervised algorithm has been proposed,which gets both labeled and unlabeled data involved.The experimental results show that the collaborative filtering algorithm based on Gaussian mixture modelis much more robust and it can alleviate the negative effect of rating noise and improve the accuracy of prediction as well.

Key words: Collaborative filtering,Gaussian mixture model,Recommender system,Semi-supervised learning