Self-attention Mechanism Fusion for Recommendation in Heterogeneous Information Network

In recent years, lots of recommendation methods based on heterogeneous information network has been proposed. However, traditional heterogeneous information network mainly relies on the similarity of meta-path to recommend, which cannot fully mine the user’s potential preference features. We propose a heterogeneous information network model using self-attention mechanism for recommendation (HINS). In our method, local neighbor information is modeled for users and items respectively via collaborative attention mechanism. Moreover, by optimizing the relationship representation between learning nodes of multi label classification problem based on meta-path, self-attention mechanism is added to the relationship representation to make the model personalized. Finally, the two parts are unified into one model for top-N recommendation. We do evaluation experiments on four different datasets, and our model achieves the best results.