Abstract:

The problems of the multinomial finite mixture model such as initialization dependence, the boundary of the parameter space and the local optimum are solved with the minimal message length criterion in this paper. We propose a multinomial finite mixture model based algorithm and apply it to clustering the behavior data of user ratings. Besides, the probability of class belonging is used to improve the Slope One algorithm. Experimental results prove that the new clustering algorithm does a better job on clustering behavior data and the new recommendation algorithm outperforms several other recommendation algorithms.   

Key words: finite mixture model;Slope One;modelbased clustering;collaborative filtering