Multi-modal recommendation method based on mutual information and improved graph auto-encoder

The invention discloses a multi-modal recommendation method based on mutual information and an improved graph auto-encoder, and the method comprises the steps: (1) constructing an article-article co-occurrence graph and an article front k similarity graph, and carrying out the normalization of the two article graphs; (2) learning effective article modal features through an improved graph auto-encoder; (3) obtaining corresponding user modal features by aggregating modal representations of objects interacted by the user; then, spreading and aggregating modal features of the user/article on the interaction graph by using an L-layer GNN; (4) adopting mutual information constraint of two levels; (5) predicting an interaction probability between the user and the article by adopting an inner product as a recommendation basis; then, training the model by adopting a multi-task training method; and (6) carrying out recommended application by utilizing the trained model. By means of the method, the modal noise problem,