Metapath heterogeneous graph aggregation method based on feature interaction

The invention discloses a feature interaction-based meta-path heterogeneous graph aggregation method, which comprises the following steps of: mapping unique codes of a user and an article and features of the user and the article into a vector representation, and generating a high-latitude feature representation corresponding to input data; mapping to the same low-dimensional feature space; using a multi-head attention model to construct preferences of the user to different user feature and article feature combinations; constructing a meta-path between the unique code and the feature; representing each node and neighbor nodes thereof in the meta-path as a feature matrix, aggregating the feature matrix, calculating vector representation after aggregation, and obtaining vector representation of each node; updating the state representation of the current node, and obtaining the final embedded vector representation of the user and article nodes after multilayer convolution; and performing dot product operation to