User-Item Graph Representation Learning Based Recommendation Models With Review Semantics

With the development of mobile networks,online data has shown an explosive growth trend,leading to serious information overload problems.Taking e-commerce platforms as an example,it is difficult for users to quickly and accurately find the products they are interested in among the massive amount of products.The real-world problem highlights the need for e-commerce recommendation-related research.Researchers have proposed a large number of recommendation algorithms,but existing recommendation algorithms still face the following two problems:(1)Some models do not explore the role of review text as an aid to interaction learning and ignore the influence of higher-order features in the user-item interaction graph.Also,they do not consider the influence of homogeneous nodes in the interaction heterogeneous graph in information dissemination.(2)The unbalanced distribution of interaction data in the user-item interaction graph and the weak supervisory signal on the graph do not allow for adequate learning of the underlying features of the graph structure using only the observed interaction information.To address the above problems,the main research contents and results of this paper are as follows.(1)In order to fully exploit the information of user-goods interaction graph structure,this paper proposes the homogeneous graph attention network recommendation algorithm with fused review semantics,abbreviated as HR-GAT.This method first extracts review features using convolutional neural networks and transforms them into vector representations to mine fine-grained semantic information.Then,the review representation is propagated to higher-order neighbor nodes as edge information,and the graph attention network is used to learn the influence of different neighbors on the central node.Finally,using the self-attention mechanism,the homogeneous aggregated nodes and their counterparts in the graph are iteratively updated with each other to enhance the influence of homogeneous nodes in information dissemination.Experiments are conducted on five Amazon recommendation datasets,and the results show that the model achieves better recommendation results than other baseline models.These verify the effectiveness of fusing review text and user-item interaction graphs for the recommendation task.(2)To alleviate the problem of insufficient supervision information due to sparse interaction data,this paper designs a recommendation algorithm based on graph contrastive learning on the basis of HR-GAT.The method first uses edge perturbation and point discard methods for data enhancement of interaction subgraphs to obtain two different views of the same interaction subgraph.Then the different views are encoded using HR-GAT to obtain their vector representations.After that,both are mapped to the same hidden space in order to adjust the view representations.Finally,the two view representations are compared for self-supervised learning to extract the implied features on the interaction graph.Experiments are conducted on five Amazon recommendation datasets,and the results show that the model has higher recommendation accuracy compared to other benchmark models.These validate the effectiveness of graph contrastive learning-assisted self-supervised models for recommendation tasks.