Predicting Purchase Intent from E-Commerce Behavior Sequences

This paper investigates the problem of predicting purchase intent based on early-stage user interaction sequences in e-commerce browsing sessions. The task is formulated as a binary classification problem, aiming to determine whether a purchase will occur using only the first three events of each session. Several modeling approaches are compared, including logistic regression, random forest, multilayer perceptron (MLP), and the Neural Attentive Recommendation Machine (NARM)—a deep sequential model that integrates gated recurrent units with attention mechanisms. A publicly available dataset from a multi-category online retailer is used to extract both aggregated session-level features and item-level behavioral sequences. Evaluation results show that NARM achieves the highest AUC (0.867) and F1 score (0.725), outperforming classical models even with truncated input. Interpretability is supported through feature importance analysis in classical models and attention heatmaps in NARM, revealing how different user behaviors contribute to predictions. These results underscore the effectiveness of sequence-aware modeling for real-time purchase intent prediction and demonstrate the complementary value of interpretable explanations in commercial applications.