HACE-Rec: Heterogeneous Graph Attention-based User-Content Interaction Feature Extraction and Personalized Recommendation

This paper proposes a content-side feature extraction and personalized recommendation method, which is based on heterogeneous graph attention model. First, the model will extract and convert the common attribute features of the content side into separate entity nodes, and perform graph modeling with the content nodes, and at the same time multi-type interaction between users and content is used to construct heterogeneous graphs to maximize the retention of information contained in the complex relationship between the user side and the content side; then, the heterogeneous graphs are used for feature mining and information analysis, which is based on the attention mechanism of graph neural network. Transmission and aggregation and update of node representation; then, fusion of information disseminated by each node on the user side and content side, calculation of matching score and integration of pairwise loss and graph loss to train the model and finally output the ranking of Top-K recommendation results. Finally, through comparative experiments with multiple benchmark models and ablation experiments on the loss function. Results show that the HACE-Rec model has improved the recall and accuracy.