Research Of Session-Based Recommendation With Graph Neural Network

Recommendation system makes use of interactive information between users and projects to recommend,which is widely used in various e-commerce platforms and media streaming websites.As an important branch of recommendation system research,the session-based recommendation system generates personalized recommendations for users by analyzing their historical behavior information.For anonymous users or new users of shopping websites,the long-term interests and short-term preferences of users can be modeled according to the historical behavior sequence data of users when their historical interaction information is unknown.By mining the user’s interest preferences,it can predict the next item that the user is likely to click on.The main work of this paper includes:In view of the problem that the users’ shopping purposes or users’ interests are mined incorrectly,and the users’ interests extraction method is simplified,this study proposes a hybrid attention mechanism of linear combination of multi-heads attention and soft attention,and then proposes a session-based recommendation method based on multi-dimensional interest extraction.The proposed mixed attention mechanism can extract users’ interests and preferences from multiple dimensions,and can directly obtain the the similarity between any two items without considering the distance between items in the session graph,so as to obtain multi-level representation of the session and reduce the influence of irrelevant behaviors on users’ preferences.In view of the decline of recommendation performance caused by insufficient consideration of the relative importance between items or information characteristics of some nodes in the session sequence,this study proposes a recommendation method based on dual-gated graph neural network.The existing session graph lacks the clicking order of items in the original session sequence,which makes the relative importance of items cannot be accurately measured.Furthermore,the item node information in the session graph may be lost after the graph neural network is updated several times.The dual-gated graph neural network proposed in this paper can capture the transformation relationship between items and obtain the correlation between nodes in the session graph and their neighbors,so as to reduce the noise interference of unrelated neighbor nodes and improve the recommendation performance effectively.The experimental research focuses on the session-based recommendation method based on multi-dimensional interest extraction and the recommendation method based on dual-gated graph neural network.Experiments were carried out to verify the mixed attention of the session recommendation method based on multi-dimensional interest extraction,it shows that this method can deeply mine the user’s shopping purpose or users’ interestsand preferenceshidden in the session,and it also can extract users’ interestsand preferences from multiple dimensions.For example,the P@20 value of MDIE under the Dgeca dataset improved by 11.95% compared with the existing model.In case of considering the relative importance of items and the information characteristics of some nodes in the sequence,the recommendation performance of the recommendation method based on the dual-gated graph neural network was verified by experiments.For example,experimental results show that the P@20 value of DGG-GNN under the Tmall dataset is 6.57% higher than the existing model.