Utilize Heterogeneous Data For Recommendation

Nowadays,internet has become the necessary facility of people's life,which leads over-whelming various personal data to be recorded such as personal profile,shopping history,rating and review data,browsing history of different kinds of media and so on.How to utilize these heterogenous data to recommend is a very important research,because these heterogenous data often contains people's different aspects of preferences or recommended items' different aspect-s of characteristics,which are definitely able to improve the performance of recommendation.There are some hard point when designing this kind of recommender system,for instance,how to extract different features from heterogenous data reasonably,how to model data with differ-ent structure,how to balance the impact of heterogenous data on the final recommender system.This paper introduces two different recommender systems which solve the above problems in two different ways.The first one is the CBPF model,which utilizes the poisson distribution to model two kinds of data:continuous rating data and discrete price data,because CBPF supposes these two kinds of data share the same distribution characteristics.The second one is the CHDR model,which utilizes a kind of special neural network autoencoder to extract features from vari-ous heterogenous data.CHDR combines the extacted features tightly with the classical pairwise collaborative model to recommendate.Experiments on several real datasets prove the value of heterogenous data and the rationality and advantages of these two models.