Sequence Recommendation Based On Graph Neural Network

Different from traditional collaborative filtering and content-based recommendation,sequential recommendation can well capture the interaction between users and items and user preferences over time by modeling the sequential behavior between users and items.Sessions are more fine-grained divisions of sequences,and session recommendation can better capture users' short-term interests;graph neural networks can vectorize non-Euclidean data,and in recommender systems,they can well capture user items.The relationship between the two has attracted more and more attention in the industry.After studying a large number of related literatures,it is found that the current session-based sequence recommendation still has the following problems: First,the sequence recommendation that focuses on the next recommendation usually only performs one modeling,and cannot make good use of the information of the entire sequence;The same user has multiple sequences with the same purpose,and the relationship and connection between the sequences are not mined;third,most sequence recommendations only focus on a single behavior between the user and the item,and do not explore the impact of different behaviors on the recommendation.For the above problems,this thesis takes the e-commerce platform as the main application scenario,and focuses on the following research:1.Aiming at the problem that the traditional session sequence recommendation only uses one modeling,it is difficult to take into account the comprehensive information expression of the entire sequence at the same time,resulting in low recommendation accuracy.A session-based graph neural network sequence recommendation model GNN-Bi GRU-TA is proposed.The model first builds the historical conversations into a directed conversation graph,uses the graph convolutional neural network to learn the node information representation in the conversation graph,and enriches the node embedding;then uses the bidirectional gated recurrent neural network and the attention mechanism to capture the global and the user in the conversation sequence.Short-term interests,and finally generate a recommendation list.Comparative experiments on two public datasets,Yoochoose and Diginetica,show that the model improves the recommendation accuracy.2.In order to make better use of multiple sequence information of the same user and to incorporate more information representation in the model,a graph neural network sequence recommendation model MA-GNN-SR based on user multi-behavior is proposed.The graph convolutional neural network is used to aggregate items of the same user in different sequences,to mine the connections between items in different sequences,and to enrich the embedding of items.For the behavior sequence,that is,the behavior sequence in which the user performs different operations on the product,the bidirectional gate is also used.Control the recurrent network to learn and learn the user sequence behavior embedding;finally,the learned items and user behavior embeddings are spliced,and the user preference expression is generated through the soft attention mechanism,and then the prediction recommendation is performed.The comparative experiments on the public data set of JD.com JData show that the performance indicators of the model have been improved.