Research Of Feature Interactive Recommendation Algorithm Based On Graph Neural Network

With the rapid development of information technology and the explosive growth of big data,how to discover effective target data is of great significance,and the recommendation system can recommend data information for users to match their personality preferences has been more and more widely used.The recommendation algorithm based on feature interaction is one of the research hotspots of recommendation systems,and the feature interaction recommendation algorithm proposed in recent years has achieved good results,but there are still some shortcomings.On the one hand,traditional machine learning methods can achieve low-order interaction modeling by violent enumeration,but cannot effectively capture higher-order feature combinations;on the other hand,deep learningbased methods using deep neural networks to learn higher-order feature interactions are implicit,and lack some explanation of the feature interaction process.To address these problems,this study is based on graph neural networks for feature interaction modeling,which are mainly as follows:(1)An attention-aware graph matching feature interaction recommendation algorithm is proposed.The method transforms the recommendation problem into a graph matching problem by extending the traditional point-to-point modeling paradigm to a graph-to-graph.First,to capture different levels of feature interaction processes,the method models user and item information as two feature graphs and models feature interactions from both internal and external interactions.The internal interaction captures the feature interactions within a single feature graph,and the external interaction captures the node matching between two feature graphs.Secondly,attention mechanism approach is introduced to capture the importance of interactions in both feature node interaction phase and node representation fusion phase.Finally,an experimental study is conducted on two publicly available datasets,and the experimental results show that the proposed algorithm has good prediction results compared to the comparison algorithm.(2)A hierarchical two-level graph fusion feature interaction recommendation algorithm is proposed.The method addresses the deficiency that attention-aware graph matching feature interactions do not introduce contextual information,and transforms the recommendation problem into a graph classification problem.First,each feature field is regarded as a node in the feature graph,and the connected edges between nodes are regarded as the interactions between features,and meaningful feature interactions are obtained by learning to update the connections between nodes through a hierarchical graph structure.Secondly,a bi-level node and graph representation generation module is designed,which contains two levels of feature interaction and fusion process,local level interaction uses edge weights to update the representation of nodes,and global level interaction dynamically captures the importance of each feature field by compression stimulus.Further,a bilinear fusion method is designed to fuse the node information from multiple perspectives for both levels of interaction.Finally,the results are experimentally compared with several baseline models to demonstrate that the hierarchical bi-level graph fusion feature interaction recommendation algorithm has higher prediction accuracy.