Explainable AI for Market Research Analytics: Attention-Based Deep Learning Framework for Consumer Purchase Behavior Prediction with Shapley Additive Explanations

Predicting whether a customer will purchase supports targeting and segmentation in e-commerce, yet accurate models are often black boxes. This study evaluates whether a compact attention-based network for tabular customer data can match strong baselines while providing consistent feature-level explanations. We propose Explainable Deep Purchase Prediction Network (EDPP-Net), which combines categorical embeddings, dense layers, batch normalization, dropout, and feature-wise attention, and is explained post hoc with Shapley values. Experiments use a public synthetic dataset of 500,000 customers, five repeated stratified splits, validation-based tuning of all models, bootstrap confidence intervals, and paired tests. EDPP-Net reached an accuracy of 92.73 ± 0.09% and an area under the ROC curve of 0.9815 ± 0.0002, statistically indistinguishable from logistic regression and gradient boosting before and after tuning; batch normalization and attention gave no measurable gain. Permutation and Shapley importance agreed and identified purchase recency as the dominant predictor, and ranking customers by predicted probability placed 69.7% of purchasers in the top 30%. Diagnostics show, however, that nearly all predictive signal comes from this one variable, whose relation to the label is probably an artifact of the simulation. The findings therefore describe explainable models on this benchmark and should not be generalized to real consumer populations.