Research On Latent Feature Mapping Based Cross-domainrecommendation Methods

With the quick development of Internet and Web techniques,e-commerce has become a new way to conduct business activities,and more and more consumers are glad to purchase the products they need via network.Compared to traditional offline shopping means,shopping online makes consumers have more choices.However,on the other hand,information renews rapidly on the Internet,which makes consumers become more easily overwhelmed by the vast products.In order to help consumers find what they really desire and improve their shopping experience,recommender systems become indispensable in e-commerce applications.Recommender systems are the information processing system in which various kinds of data is gathered actively.They can establish the relationship between users and products,filter out the related products which suit user's taste,and finally recommend the products to the user.Collaborative Filtering(CF)is a widely adopted technique in recommender systems.Traditional CF models mainly focus on predicting a user's preference to the items in a single domain according to the ratings in the domain,such as the movie domain or the music domain.For such models,ubiquitous data sparsity problem is a major challenge.Moreover,CF cannot make personalized recommendations for the cold-start users who have no ratings at all.Recently,Cross-Domain Collaborative Filtering(CDCF)is proposed and it aims to improve the recommendation performance by taking advantage of multi-domain ratings and effectively transferring knowledge across different domains.Even so,it is still difficult for existing CDCF models to tackle the cold-start users in the target domain.In this paper,we focus on the ubiquitous cold-start user problem and propose an effective cross-domain recommendation model,Cross-Domain Latent Feature Mapping(CDLFM)model.Firstly,the cold-start users of an item domain may have ratings in another item domain.For example,in the e-commerce website Jingdong,many users tend to buy electronic products while they are much less interested in purchasing books.Secondly,the user preference in different item domains is correlated.For example,users who like comedy movies may also prefer humorous books.Therefore,the problem setting studied in this paper is that cold-start users only have ratings in the auxiliary domain while having no ratings in the target domain.Besides,there exists linked users,who have ratings in both the auxiliary domain and target domain.The goal of this paper is to take linked users as the bridge to transfer knowledge from the auxiliary domain to the target domain and make the rating prediction for cold-start users in the target domain.Based on this problem setting,we propose the CDLFM model in this paper.In the first step of CDLFM model,we handle the rating matrices in different item domains separately for learning domain-specific latent features.In order to alleviate the impact of data sparsity and provide essential knowledge for neighborhood based feature mapping,we propose a new rating matrix factorization model,Matrix Factorization by incorporating User Similarities(MFUS).MFUS studies users' rating behaviors from different point of view and incorporate the similarities between users'rating behaviors into the rating matrix factorization process.In the second step,to effectively transfer knowledge from the auxiliary domain to the target domain,we propose the neighborhood based latent feature mapping method to learn the cross-domain user latent feature mapping function.For each cold-start user in the target domain,we first find his/her neighbor linked users,with whom he/user has similar rating behaviors in the auxiliary domain.Then the latent features pairs of these linked users in both domains are used to learn the cross-domain feature mapping function which is suitable for the cold-start user.Finally,according to the learned feature mapping function and the cold-start user's latent features in the auxiliary domain,we can predict his/her characteristics in the target domain and then make the preference prediction.Our major works are summarized as follows:1.In order to alleviate the data sparsity in single domain and provide essential knowledge for neighborhood based feature mapping,an improved rating matrix factorization model MFUS is proposed.In MFUS,besides the observed ratings,the user rating behaviors are also taken into consideration and three similarity measures are proposed.As far as we know,MFUS is the first proposal to consider users' similarity relationship reflected from the sparse rating matrix in the matrix factorization process.2.The neighborhood based feature mapping method is proposed for more accurate cross-domain latent feature mapping.As far as we know,all the existing feature mapping methods for cold-start users learn one feature mapping function for all cold-start users based on all the linked users.However,this means cannot achieve personalized feature mapping,and learning based on all the linked users may introduce noise.The neighborhood based feature mapping method is not only more explicable,but also can learn more accurate feature mapping function.Besides,we adopt two mapping models,Gradient Boosting Trees and Multi-Layer Perceptron,to learn the feature mapping function and two CDLFM models,CDLFM-GBT and CDLFM-MLP,are proposed.3.We extract two datasets from the real Amazon rating data on which extensive experiments are conducted.The experimental results show that by utilizing the ratings from auxiliary domain,our CDLFM models can effectively make the rating prediction for cold-start users in the target domain.Moreover,we also experiment a lot for parameter sensitivity analysis in order to better evaluate our CDLFM model and MFUS model.