Information-aware multi-behavior processing for purchase decision prediction via cascaded graph and hypergraph learning

Purchase decision prediction in e-commerce is difficult because auxiliary behaviors are common but noisy, while purchase signals are sparse and more reliable. Existing GNN-based methods often construct bipartite graphs by treating observed interactions as true intent, which introduces noise and weakens the modeling of high-order user-group patterns. To address this problem, we propose IM-CGHL, an information-aware multi-behavior method that organizes user interactions into a three-level intent hierarchy: cognitive, affective, and conative. To reduce cross-view noise, we design a cascaded graph structure learning module that uses purchase signals to reconstruct lower-intent behavior graphs in a top-down manner. To supplement the limited information in pairwise interactions, we further introduce a fuzzy hypergraph module that links users through bridge items and captures high-order group semantics within a shallow network. Experiments on three real-world e-commerce datasets show that IM-CGHL consistently outperforms strong baselines, achieving F1-score gains of 6.87% on JingDong and 11.28% on TaoBao over the state-of-the-art baseline MBGCN. These results show the value of combining structure denoising with group-level information enhancement for purchase decision prediction.