Research On Review-based Cross-domain Recommendation

In recent years,the vigorous development of online product review websites has brought great convenience to users,and at the same time,users also leave a lot of pref-erences information on them.Through this information,websites get users’ interests and recommend products to users.Thus,recommendation systems(RSs)come into being.RSs predict products a user may like for user’s reference by calculating preference in-formation.This has been widely used in product review websites,such as Amazon,eBay,Taobao,JD.com,etc.,and has achieved great success.However,until now,the new user’s cold start problem is still a challenge which cannot be ignored in RSs.As an effective solution,cross-domain recommendation(CDR)is attracting more and more attention.CDR commits to take advantage of knowledge transfer,transferring users’ preference from a relatively dense domain to a relatively sparse domain to assist in resolving cold start problem in target domain.Most of the existing CDR works exploit ratings only,but in fact,text such as review text con-tains richer user preference,which is usually overlooked.In review text,users usually express emotion preferences in fine-grain level,using these will greatly help improve the accuracy of CDR.For example,for a cell phone,a user’s review can show us his opinion on "screen" or "battery",which is much more accurate than using only the rat-ings.Based on the above,this paper studies how to effectively use the fine-grained pref-erence from review text for CDR.Since the preference information in review text is disorganized,the use of fine-grained preferences is more complicated than the use of ratings only.There are two challenges:first,how to model users’ fine-grained prefer-ences from review text to adapt to the needs of knowledge transfer?And second,how to effectively transfer the fine-grained preferences?In this paper,a relatively in-depth study about this is done.This paper proposes a relatively complete solution framework, containing a series of relatively complex issues.The specific work and contributions are summarized as follows:1.For the modeling of fine-grained preferences in cross-domain frameworks,this paper introduces the "aspect" level into the framework,and uses "aspect rating tensor"and "concern degree tensor" to model user preferences.Ternary relationship of "us-er-item-aspect" is used to organize fine-grained preferences extracting from review text,and third-order tensor("tensor" for short in below)is exploited to model the ternary relationship.In a tensor,each element records the preference of a user in an "aspect" of an item.The advantage of this modeling approach in cross-domain recommendation is:first,the disorganized user preferences in review text can be organized into structured data that facilitates the alignment of two domains;second,taking "aspect" as an inde-pendent dimension can facilitate to capture the relation of two "aspects" from different domains,so that the role of "aspect" level is fully exploited in cross-domain recom-mendation.After introducing the "aspect" level,in order to solve the problem of the re-lationship between the preference of a user in each aspect of an item and the overall preference the user to the item,based on the idea of "the ratings on the aspects which gain more attention usually play a more significant role in the overall rating so that they’d better be allocated greater weight",this paper models "aspect rating tensor" and"concern degree tensor" in each domain.2.For the problem of fine-grained preference cross-domain transfer,this paper proposes the Review-Based Joint Tensor Factorization(RB-JTF)model,and applies it to the cross-domain recommendation framework.The RB-JTF model transfers knowledge by jointly factorizing two tensors from different domains.Until now,how to transfer knowledge through jointly factorizing two tensors from different domains is a big challenge.In RB-JTF,first,two tensors are decomposed into three factor matrices,which are the latent factor matrices of users,items and "aspects".And then,the RB-JTF transfers knowledge,including two parts:first,share the users’ latent factors among auxiliary domain and target domain;second,capture the relations of two "aspects" from different domains and assimilate them into the joint tensor factorization model,in which the semantic similarity of two "aspects" is first measured,and then varying degrees of restraint are given to the distance of two "aspects" in latent factor space based on the semantic similarity.Achieve the transfer of "aspect" latent factors.This paper gives the objective function of RB-JTF model,and the factor matrix is learned by nonlinear con-jugate gradient method,and the time complexity of the algorithm is analyzed.Different from the previous cross-domain recommendation model,RB-JTF predicts the prefer-ences on "aspect" level.In order to predict user’s overall ratings,knowledge transfer applies on both "aspect rating tensor" and "concern degree tensor".In addition,for the sparse problem of tensors,this paper proposes a mitigation stategy.By pre-estimating some of the elements in tensors and their reliability,the accuracy of joint tensor factori-zation model is further improved.3.This paper validates the effectiveness of the proposed framework on the Ama-zon.com dataset.In this paper,two pairs of domain are used,which are "Movies-Books" and "Movies-Music CDs".Experiments show that the RB-JTF model is supe-rior to the baselines in RMSE and MAE,especially for the cold start users in the target domain.In addition,the experiments also carry out the impact of important parameters analysis,learning algorithm iteration analysis and performance analysis of the mitiga-tion strategy of sparse problem in tensors.