Research And Implement Of Recommender Algorithm With Side Information

Recommender system has brought convenience to people's life continuously since it was put forward in the 1990 s,and the current information age is a great time for recommender algorithms to work.Recommender system is of great significance for ecommerce platforms,so that it can help platforms to increase traffics,earn advertising revenue and increase sales of goods,etc.Therefore each major e-commerce platform has its own recommender system basically,and these platforms are flooded with available information.It has become a trend to use various side information to improve the prediction performance for the recommender systems.In line with this trend,this thesis focuses on how to better combine side information to improve the performance of recommender algorithms in e-commerce platforms.However,with the increase of ecommerce platforms,users' browsing difficulty is also increasing,and users often need to switch to different websites for comparative shopping,which is also inconvenient for users,therefore it has practical significance for the research and development of crossplatform e-commerce websites.The main work of this thesis includes the following:(1)A fusion recommender model named HRNrec is proposed in this thesis,which obtains the feature relationship between users and items through two generating graph methods,and then inputs it into the fusion neural network model to generate recommendation results.The first generating graph method is to construct a tripartite graph of user-item-category,then carry out the walk probability calculation through the random walk algorithm,and finally generate the first part of the input of the fusion recommender model.The second generating graph method is to use heterogeneous information networks to obtain the meta-graph,solve the similarity matrix of users and items in the meta-graph through the meta-path calculation method,and then obtain the relevant feature vectors through the probability matrix decomposition method,so as to generate the second part of the input of the fusion recommender model.Finally,the performance of the algorithm is verified by relevant experiments.(2)This thesis proposes an improved graph-based recommender algorithm named MoTSPR for e-commerce websites,which contains a weight adjustment strategy and a probability adjustment strategy,and this algorithm can be applied to repetitive recommendation scenarios,finally the effectiveness of this algorithm is verified on realworld datasets.(3)This thesis designs and implements a cross-platform e-commerce recommendation website named Bibikan,which provides users with such functions as commodity price comparison,historical price inspection,price reduction reminder,and personalized recommendation service.At the same time,the MoTSPR algorithm proposed in(2)is applied to Bibikan website.