Research On Personalizing Session-based Recommendation System With Dual Encoders

With the development of big data and deep learning technology,intelligent recommendation systems have penetrated into all aspects of people's lives.When a user is visiting a website,the server of the website usually records the user's access record in a period of time in the form of a session.A recommendation system built for session is usually called a Session-based Recommendation System,which is a branch of the recommendation system.The purpose of session recommendation is to predict the item that the user may be interested in at the next step based on the user's previous click sequence,so as to provide the user with better services.In the past,sessions were usually regarded as sequences,so researchers often use a recurrent neural network to model every session.Therefore,this kind of models are usually unable to fully capture the complex transitions of items.Moreover,previous studies generally believed that all sessions were anonymous,so each session was regarded as independent.Thus,previous studies ignored the fact that different user have different interest on items.In order to improve the two deficiency above,we propose a personalized session-based recommendation model with dual encoders,PSRDE for brevity.The PSRDE model adopts the encoder-decoder structure.The encoders contain a session encoder and a user encoder.The session encoder can model current session and the user encoder can model current user's historical session record.In the decoder,the multiple features output by the encoder are fused to obtain the final feature vector,and then a list of recommended items can be obtained through calculation.PSRDE uses the gated graph neural network which can capture more transitions of items to model every session,and PSRDE not only uses the current session,but also takes the user's personal preference into account,so that PSRDE can achieve better recommendation effects.In order to verify the validity of the model,extensive experiments conducted on two real datasets show that PSRDE model outperforms several baseline session-based recommendation methods.