Hypergraph Neural Network Algorithm for Complex Market Relationship Modeling and Consumer Behavior Prediction

Abstract

High-level complex relationship systems that drive consumer behavior exist among today's market consumers. Traditional models (e.g., GNNs) treat the market as two-party graphs, which limits their ability to replicate complex dynamics, such as group purchasing, brand loyalty among multiple firms, or interchangeability of products across multiple categories. This paper presents the HyperGNN-Market model with an approach to developing a hypergraph neural network (HGNN) for modeling the greater order relationships found in markets, thus improving the predictive ability of consumer behavior by developing a hypergraph of relationships between market participants (consumers and retailers), their products, and the environmental cues connecting them. The hyperedges of the hypergraph represent the n-party relationships among the participants, their products, and environmental cues; for example, hyperedges could represent groups of consumers that purchase together, purchase in response to promotions, purchase from common demographic groups, or purchase products that are substitutable. A new dual-channel mechanism is presented for refining node features in hypergraphs called Dual-Channel Hypergraph Attention Propagation (DHAP). This approach combines both local attentional refinement per hyperedge as well as global attentional refinement across hyperedges. Thus, DHAP can capture individual/group level behavior (i.e., intra-group cohesiveness) and market segmentation behaviors (macro) simultaneously.  The performance of HyperGNN-Market is evaluated against eleven baseline methods across four large-scale datasets in predicting purchases (i.e., received +9.4% improvement in AUC), recommending next basket (i.e., received +14.2% improvement for NDCG@10), predicting customer churn (i.e., received +11.7% improvement on F1 score), and forecasting market share (i.e., received −14.8% improvement in MAPE). In addition, comprehensive ablation studies demonstrate the contribution of each architecture component, and a deployment with an unnamed major Southeast Asian eCommerce company provided evidence of a 6.3% increase in conversion rate.