A weighted heterogeneous graph attention network method for purchase prediction of potential consumers with Multibehaviors

In the e-commerce environment, predicting the purchasing intention of potential consumers is an important component of recommendation systems, which provides a basis for personalized recommendations by predicting whether users are likely to purchase a certain product. This accurate prediction not only enables businesses to cater to consumers’ needs and preferences, thereby stimulating purchases, but also guides promotion and advertising efforts. However, most current research uses a single data structure, which may have certain limitations in improving prediction accuracy. Therefore, in order to construct a more effective purchase prediction method, this study constructs a weighted heterogeneous graph attention method based on various interaction behaviors between users and products. Firstly, we construct a multi-behavior bipartite graph based on user–product interaction behavior. Next, the user–product multi-behavior bipartite graph is reconstructed into user relationship graph and user–product relationship graph. Then, we use multi-head graph attention network to learn the neighbor node information in user relationship graph and user–product relationship graph respectively. Finally, we utilize a linear attention mechanism to automatically learn the importance of different relationship graphs in predicting user purchase intention. The effectiveness and superiority of our method is confirmed by the comparative and ablation studies conducted on the dataset of potential consumer purchase behavior provided by JD.com. Specifically, after training and parameter optimization, our method is able to achieve a precision of 0.965, a recall of 0.974, and an f1-score of 0.969, which all outperform the comparison methods.