Category Attention Graph Neural Networks For Session-based Recommendation Research And Implementation

Recommendation systems are widely used in various fields of life.It does not require users to provide accurate data input,and can recommend information of interest to them,which can effectively solve the problem of Internet information overload.Conversational recommender system is an important branch of recommender system application field,which is generally used in anonymous user recommendation.During the operation of the recommendation system,the personal information of anonymous users cannot be normally obtained by the recommendation system.The conversational recommendation system can make accurate personalized recommendations for the user by mining the implicit user interest information in the anonymous user's conversation sequence when the user's personal information is missing.Existing studies have shown that traditional conversational recommendation systems exhibit hysteresis in multiple rounds of conversations,and specific weight settings in user preference mining and identification will lead to a high deviation rate.Recent studies have shown that graph neural networks can model traditional conversation sequences as directed conversation graph structures,thereby effectively capturing the information flow in multi-round conversation sequences.Various conversational recommendation algorithms based on graph neural networks have become the frontier of research.Through research and experimentation of existing mainstream methods,many of them have the following shortcomings:(1)The existing mainstream conversational recommendation algorithms ignore the important item feature of item category,and the algorithm cannot learn the deep feature mapping between item category and user interest,thus limiting the recommendation performance of the algorithm;(2)Most of the existing mainstream conversational recommendation algorithms use common public datasets,and the recommendation effect on datasets based on other application fields has not been verified,thus reducing the generality of the algorithm;(3)Most of the related algorithms are only in the laboratory research stage,and there are very few recommendation systems applied to the real application platform environment,and the actual recommendation effect of the algorithm cannot be verified.In view of the above problems,this paper has completed the research on the conversation recommendation algorithm based on graph neural network and related experimental design.The specific research is as follows:(1)Research and design a session recommendation algorithm based on category attention graph neural network(Category Attention Graph Neural Networks for Session-based Recommendation,CAGNN).The model uses a graph neural network to model session sequences,and learns the state transition between session clicks through a Gated Recurrent Unit(GRU),and studies category attention to enhance the capture of different The interest of the category click item is transferred,and finally the embedded session vector generated by learning is trained and the next click item of the session is predicted,which effectively improves the recommendation performance of the algorithm.(2)Research and design comparative experiments in multiple dimensions,and verify,analyze and improve the recommendation performance of the neural network model based on the category attention map through experimental data.Based on the session recommendation algorithm described in this paper,a real-world user behavior dataset(XCF2021)is customized,and this dataset is used to extend the generality and robustness of the algorithm in this paper.At the same time,a control experiment was designed to compare the algorithm model in this paper with a number of common conversation recommendation methods,to verify the actual performance of the CAGNN model and improve it;through the design of ablation experiments to explore the impact of the four different network structures on the performance of the CAGNN model.Find the best neural network configuration;design long and short session experiments to analyze the recommendation performance of four deep learning algorithm-based models under different session lengths.Finally,the influence of different implicit vector dimensions on the recommendation performance of the four models based on deep learning algorithms is analyzed by studying the implicit vector dimension and conducting related experiments.Through the above experiments,the generality of the algorithm is fully expanded and the robustness of the algorithm is improved.(3)Through the above research,design and implement a conversational recommendation algorithm,and build a corresponding recommendation system.The system uses real platform data in the field of gourmet cooking to model and fine-tune the corresponding model,and finally trains it into a commercial real-world recommendation model.The model uses the session recommendation algorithm based on the category attention map neural network to make corresponding recommendations for users,thereby verifying The recommendation effect of the CAGNN model on the actual platform is shown.