| When facing massive amounts of e-commerce data,recommendation system can provide users with interest information through analyzing users’ historical behaviors.Though information overload can be relieved to a certain extent based on some existed recommendation methods,the accuracy and universality of recommendation systems are still unsatisfactory when facing the cold start problem caused by the lack of interaction information from new users.From the above analysis,the paper will conduct in-depth researches on how to improve the accuracy and universality of recommendation system.(1)To improve the richness of interaction information,an auxiliary classification network based text sentiment classification model is proposed,and user comment text sentiment can be seen as another important recommendation basis besides score.The model constructs a new auxiliary classification network,which enlarge the differentiation between the text features with different sentiments;In addition,the model also adopts a weighted domain discrimination network to align the text features with same sentiments in different domains.Experimental results show that the proposed model can obtain high text sentiment classification accuracy in target domain,and its architecture has important reference significance for constructing the following aspect-level sentiment classification model;(2)Since the overall sentiment of comment text can not reflect user’s preference and the unique attributes of commodities well,a multiple adversarial network model with inter-class is constructed,which achieves cross-domain comment text aspect-level sentiment classification.The model adopts an aspect-level domain discrimination network to achieve the weighted alignment of the comment text features with multiple aspect-level words.In addition,the feature representation ability has been improved through restricting the distribution difference of text features between different categories.Experimental results show that the proposed model can achieve high aspect-level sentiment classification accuracy in the target domain,which provides important data support for the subsequent interest recommendation;(3)To verify that comment text is helpful for improving recommendation performance,a non-negative matrix decomposition model based on comment and score fusion is established,which achieves commodity recommendation under the same domain.First,the recommendation evaluation matrix is constructed by the weighted fusion of text feature and score feature;Then,when decomposing the evaluation matrix,the reconstruction constraint,the similarity constraint and the sparseness constraint are all imposed to ensure the accuracy,the invariance,and the interpretability respectively.Experimental results show that compared to use score feature only,the recommendation performance can be improved effectively after fusing text sentiment feature.(4)To improve the transferability of recommendation knowledge between different domains,a score and text fusion based multi-attention model is constructed,which achieves cross-domain commodity recommendation.First,the domain discrimination network is used to align the shared hidden features and the comment text features between domains,and the transfer bridge of recommendation knowledge can be established;Then,the unique and shared hidden features are fused based on the proposed attention mechanism,and the real interest can be acquired.Experimental results show that even if lacking of enough interactions in the target domain,the cross-domain interest recommendation for users can still be achieved well.In summary,the research theme and contents of this paper are of important theoretical research value.At the same time,the research work has good reference significance for optimizing resource allocation and adjusting marketing strategies for businesses,and brings potential significant commercial value to company.This paper has 44 figures,37 tables and 177 references. |
PDF Full Text Request