Incorporating Side Information With Graphic Models For Recommendation

With the popularity of the Internet and the rapid development of computer technology,more and more data swarm into our lives.Massive data has caused the problem of information overload and users are easily disturbed by useless information.Recommendation systems can effectively alleviate this situation by analyzing user's historical behavior,studying user's interest preference,and actively recommending the information that may be interesting to the user.Traditional recommendation algorithms recommend by using user's historical ratings.But due to the sparse rating matrix,it is difficult to generate appropriate recommendation for the user who lacks data.However,in addition to ratings,there are many side information can help to mine user's preferences.Information about user,item,or behavioral environment may affect the user's preference for the item.As a method of intuitively describing the relationship between variables,graphic model is being applied more and more to recommendation systems.Algorithms based on graphic models can flexibly introduce new variables and model the interrelationships between variables.Based on the graphic model algorithm,this paper studies how to improve the recommendation performance by using side information other than ratings.In this paper,we propose a preference-aware recommendation algorithm based on matrix factorization which incorporates categorical information.At present,most matrix factorization based recommendation algorithms represent systematic tendencies of user's rating behaviors by introducing the bias term which can not capture the relationship between users and items implied in ratings.In this paper,the preferenceaware recommendation algorithm we proposed incorporates categorical information to capture the preferences between users and items under the framework of matrix factorization.The Gaussian prior distribution is used to generate user latent feature vector and movie latent feature vector.According to the different types of categorical information,user latent feature vectors and movie latent feature vectors are generated in different ways.One is generated by a separate Gaussian distribution,the other is generated by a mixture of Gaussian distribution.The algorithm improves the accuracy of rating prediction by adding a preference factor.Experimental results on the two MovieLens datasets show that the preference-aware recommendation algorithm incorporating categorical information that we proposed can effectively use categorical information to improve the recommendation performance.Related research results have been published in The 15 th IEEE International Conference on Machine Learning and Applications(ICMLA)which was held in Anaheim in December 2016.This paper also proposes a recommendation algorithm based on topic model which incorporates text information.Traditional recommendation algorithms can not use text information to recommend.To this end,this paper proposes a recommendation algorithm based on topic model,which extracts movie's topics and user's interests from the data containing the text information,and gives recommendations according to the correlation between the distribution of topics of movies and the distribution of user's interests.The proposed algorithm based on the R-LDA model can handle user's rating data and text description information of movies at the same time.The R-LDA model associates user's interests with the text description information of movies while modeling topics of movies,mining the relationship between user's ratings and user's interests plus topics of movies.Experimental results show that the R-LDA recommendation algorithm incorporating text information can effectively improve the recommendation performance in top-N recommendation task.