Research On Multi-interest Sequential Recommendation Method Based On Neural Network

In recent years,recommendation systems have become indispensable tools in social networking sites,e-commerce,video and entertainment,and other fields.The sequential recommendation is taken a series of "user behavior interactions" as input,tried to model the complex sequence dependencies embedded in user-item to predict the subsequent user-item interactions that may occur soon,generated a recommendation list consisting of the highest ranked candidates by using the maximization of utility functions.The purpose of the multi-interest sequential recommendation is to generate multiple vectors for users,which can obtain users’ interests more accurately.According to data sparsity problem of the existing multi-interest sequential recommendation,a multi-interest sequential recommendation model MIGAN＿Rec is put forward based on a generative adversarial network.In the MIGAN＿Rec model,it is proposed to obtain the prediction sequence through user behavior sequence extension layer.In the user behavior sequence extension layer,the generator is used to predict the user’s next click,and the discriminator is used to determine whether the recommendation sequence generated by the generator is true.The experiments on the MIGAN＿Rec model are been on Amazon Books and Taobao datasets.It is seen that the MIGAN＿Rec model can achieve better results than other models in the multi-interest sequential recommendation.However,the improvement of performance limited in the improvement of accuracy in the complex data type of multi-interest sequential recommendation.Although the problem of data sparsity is solved in the MIGAN＿Rec model to a certain extent,it can’t obtain user dynamic preferences.To obtain users’ dynamic preferences,a multi-interest sequential recommendation model MIGANGCN＿Rec based on a neural network is proposed.On the basis of MIGAN＿Rec model,a user dynamic preference acquisition layer is proposed and the prediction sequence obtained in the user sequence extension layer is input into the user dynamic preference acquisition layer to obtain the item representation vector.In the user dynamic preference acquisition layer,each item node and its corresponding neighbor node are iteratively represented by using the graph convolutional network.The experiments in this paper are conducted on the MIGANGCN＿Rec model on Amazon Books and Taobao datasets,and the results show that the MIGANGCN＿Rec model can achieve better results than other models in getting users’ dynamic complex preferences.and the improvement of the experimental indicators is very significant on Taobao dataset with multiple relationships.The proposed MIGAN＿Rec model in this paper is adopted a generative adversarial network to eliminate noise,which effectively addressed the challenge of large-scale sparse multi-interest sequential recommendation.On this basis,the model MIGANGCN＿Rec is proposed to enhance the effect of multi-interest sequential recommendation.