In order to solve the dual dynamic coupling modeling problem in `social-consumption' interactions in S-Ecom, this paper presents a Dynamic Multi-Task Graph Neural Network model (i.e., DM-GNN) which co-predicts evolving patterns of user social relations and consumption behaviors. First, we build a temporal heterogeneous graph: social interactions (e.g., likes, comments, and reposts) among users are modeled as timestamped edges, and consumption records (e.g., browsing, adding to cart, payment) are represented as dynamic node attributes. Second, we design a dynamic weight decay strategy to incorporate an exponential time decay factor into message passing for measuring the “recency effect” of social influence diffusion at each time stamp. Finally, we adopt a Temporal Graph Attention (TGAT) network as the shared encoder that can leverage a multi-task learning framework to jointly optimize link prediction (social relationships) and multi-class classification consumption tiers in an interacting way, which helps mutual enhancement of social structures and consumption patterns. Experiments on the AliGraph 54 Tmall subset (12,890 users, 457K social edges and 1.2M consumption records) show that DM-GNN outperforms compared approaches: for social relationship prediction model, it enjoys the AUC up to 0.843 which is superior than our best baseline TGAT by 4.1%; and for consumption behavior prediction task, it reaches Macro F1 score as high as 0.689 which significantly exceeds single-task learning methods as well as static graph based approachs. The ablation results also validate that the dynamic decay mechanism and multi-task cooperation make a significant contribution to the performance. Model interpretability analysis also uncovers about the positive promotional effects of high-influence consumers on their social ties'consumption, offering actionable decision support for phenomena such as live-streaming commerce and community-based marketing.
Zhang et al. (Sun,) studied this question.