Robust prediction of shopper intent via heterogeneous feature interaction and gradient calibration

Online purchase intention prediction using clickstream data is constrained by the inherent heterogeneity of tabular features and severe class imbalance (up to 1:9). Standard deep learning architectures often implement uniform processing for categorical and continuous variables, failing to account for high-order feature couplings and the gradient dominance of the majority class during optimization. This paper proposes an Adaptive Heterogeneous Attention Network (AHAN) to address these structural limitations.The architecture incorporates two primary components: (1) a Multi-Branch Contextual Attention Module (MBA-CM) that projects features into parallel subspaces to isolate and model high-order nonlinear interactions between numerical and categorical inputs; and (2) an Adaptive Gradient Focusing Loss (AGFL) that reweights the loss function to prioritize minority-class samples during backpropagation. The proposed framework was validated on two distinct datasets: the public OSPI benchmark and a real-world global e-commerce behavior (GEB) dataset. Empirical results demonstrate that AHAN achieves a Precision-Recall AUC (PR-AUC) of 0.702 on OSPI and 0.638 on GEB, representing a statistically significant improvement over Gradient Boosting Decision Tree (GBDT) baselines and standard deep tabular models such as FT-Transformer. The findings confirm that type-sensitive feature interaction and adaptive gradient modulation effectively improve predictive robustness across both public and production-scale imbalanced datasets.