Research On Deeply Fusing Reviews And Contents Based Cross-domain Recommendation Methods

With the rapid development of e-commerce,the number of purchasing choices for customers is growing rapidly.Recommender systems have become an essential tool to help customers find the commodities that suit their tastes better.A preeminent recommender system is able to help a company cater to customer preference so that it can assist in retaining users and increasing sales.However,for a single user,the products he/she ever bought or used only account for a very small part of all the products on the shopping website in that he/she is involved in very limited information in a particular domain,which leads to the sparsity of the rating matrix.If a user is new in a domain,we are not able to acquire his/her history information,so that it's difficult to make satisfying personalized recommendations for the user.This situation is called as the "cold start problem"of new users.If we can solve or alleviate cold start problem in some way,we are able to improve the precision of recommendation to some extent and improve the service level of enterprises.Hereto,cross-domain recommendation has gained increasing research interest recently and it offers a new perspective to slove this problem.Cross-domain recommendation aims to better acquire users'personalized preference by means of transferring explicit or implicit feedback from the auxiliary domain with dense information to the target domain with sparse information.Unfortunately,even though existing cross-domain recommendation algorithms have achieved fine accuracy of recommendations,most of these algorithms only utilize users'ratings which is limited to the inherent sparsity of rating matrices.Although some single domain recommendation works have shown that side information such as review texts,item contents and tags could contribute to improving the recommendation performance,most existing cross-domain recommendation works still do not take full consideration of different kinds of valuable side information and cannot fully and deeply fuse the side information with the rating matrix.Based on this,in this paper,in order to improve the performance of cross-domain recommendation,we study how to take different kinds of side information into consideration and make them cooperate with ratings in a deep fusion way.The challenges are as follows.First,which kinds of side information are suitable for deeply fusing with rating matrices?Second,how to fuse these side information with rating matrices?Hereto,in this paper,we deeply studied and proposed a Review and Content based Deep Fusion Model,namely RC-DFM,to alleviate data sparsity and make better personalized recommendations for cold start users.There are four steps of RC-DFM.Step one,vectorization of reviews and contents.In this paper,we choose user reviews and item contents as side information.Before we apply these side information on our model,we need to transfer them into their corresponding vectors with semantic information so that they can be handled by neural networks.RC-DFM use word-embedding and max-pooling to transfer the reviews each user ever wrote,the reviews each item ever received and the contents of each item in auxiliary domain into their corresponding fixed vectors with semantic information.Step two,generation of user and item latent factors.We need to better describe the latent features of users and items to improve the performance of recommendation systems.RC-DFM jointly trains several extended versions of stacked denoising autoencoders(SDAE)to deeply fuse these text information and rating matrices in an effective way.In this way,we can acquire the latent features of users and items with rich semantic information.Step three,cross-domain nonlinear mapping of latent factors.We build the nonlinear mapping relationships between the users in auxiliary domain and target domain by training a multilayer perceptron(MLP).Step four,cross-domain recommendation.Based on the nonlinear mapping function learned from MLP and the extracted latent factors of cold start users in auxiliary domain,we can get the latent factors in target domain and make personalized recommendations according to these.Our major contributions are summarized as follows:1.We design a kind of new framework to learn the latent factors of users and items in single domain using a deep learning method.We extend several basic SDAEs and jointly train them so that we can deeply fuse rating matrices,user reviews and item contents in both auxiliary and target domains.In this way,the latent factors we learned can be full of semantic information and also own the information mining form rating matrices.2.In order to make better personalized recommendations for cold start users in cross-domain recommender systems,we propose a Review and Content based Deep Fusion Model,namely RC-DFM.This model can use deep learning to fuse different kinds of information to increase the density of data,and it utilizes cross-domain nonlinear mapping to make satisfying personalized recommendations for cold start users who have no history information at all in target domain.3.We choose the public Amazon dataset to verify the superiority of RC-DFM.We choose two pairs of domains,"Movies&Books"and "Movies&Music CDs"to perform the experiments.Based on the results,we can observe that RC-DFM acquires better performance both in terms of RMSE and MAE compared with the baselines.Moreover,we conduct extra fine-grained experiments to verify the effectiveness of the addition of review texts and item contents for improving the performance of recommendation systems.