PinGAT: Hybrid GNNs for Customer Classification and Product Recommendation in Online Retail

Abstract

Large-scale e-commerce transactional networks are often affected by revenue skewness, graph sparsity, evaluation leakage, potential class imbalance, and limited model transparency. This paper introduces PinGAT, a hybrid Graph Neural Network (GNN) framework for customer classification and personalized recommendation that integrates the scalable neighborhood aggregation of PinSage with the localized attention mechanism of Graph Attention Networks (GAT). For supervised settings with imbalanced class labels, PinGAT supports trainingonly SMOTE augmentation and feature-space k-NN structural matching. In the empirical setting, because the benchmark ecommerce datasets do not provide predefined customer spending labels, we construct supervised spending coalitions using a temporally separated, quantile-based partitioning of future-period revenue, while prior behavioral RFM-I and spatial indicators are used as predictive features. Explainable product recommendations are then generated through an intra-cluster ranking mechanism based on bounded positive cosine similarity and a logarithmic coalition-popularity prior. Experiments on the UK Online Retail and Brazilian Olist datasets show that PinGAT achieves multiclass classification accuracies of 0.7978 and 0.7949, respectively, outperforming standalone PinSage and GAT baselines by up to 0.4208. An induced class-imbalance experiment further shows that the SMOTE-integrated PinGAT pipeline remains robust when class frequencies are skewed. The framework also produces interpretable coalition-based recommendation results, including a Brazilian Olist NDCG@10 of 0.1715 and Recall@10 of 0.2192, and an Online Retail Accuracy@10 of 0.3825. Model diagnostics show consistent alignment between PinGAT’s global feature weights and localized GNNExplainer attributions.