Temporal Heterogeneous Graph Neural Network for User Influence Prediction in Social E-Commerce

Abstract

This study proposes a user influence propagation model based on dynamic heterogeneous network representation learning. The model combines multi-type nodes and semantic edge types, and adopts a learnable temporal embedding strategy to construct a dynamic heterogeneous graph, so as to realize the time-aware representation of nodes. In the representation learning stage, a multi-head attention mechanism is introduced to enhance context modeling and propagation path awareness. Meanwhile, structural contrastive learning is used as an auxiliary task to improve the discriminability of node representation, and a joint training strategy of node classification and edge prediction is adopted to enhance the generalization ability of the model. In the experimental evaluation, the proposed model is compared with several representative Graph Neural Network (GNN) models, including Graph Convolutional Network (GCN), Temporal Graph Attention Network (TGAT), and Heterogeneous Graph Transformer (HGT). In the overall prediction task, the proposed model achieves a Mean Squared Error (MSE) of 0.1214, a Mean Absolute Error (MAE) of 0.2398, and an R² of 0.701. All these metrics are significantly better than those of TGAT (MSE: 0.1597, MAE: 0.2785, R²: 0.603) and HGT (MSE: 0.1429, MAE: 0.2652, R²: 0.642). Within a 14-day prediction window, the model maintains an error rate of 0.1390, demonstrating superior temporal generalization capability. For identifying high-influence user groups, the model achieves an MSE of 0.1268, significantly better than HGT’s 0.1456, indicating higher sensitivity in modeling strong propagation nodes. This shows that the combination of these methods enables the proposed model to demonstrate stronger temporal modeling capability and heterogeneous relationship understanding capability compared with TGAT and HGT, in terms of capturing users’ historical behaviors and predicting future influence.