Behavioral Session Segmentation and Leakage-Safe Purchase-Intent Modeling: A Methodological Framework

Background: E-commerce clickstream records can be used to describe browsing patterns and to develop session-level purchase-intent classifiers, but conclusions about predictive effectiveness require estimates obtained on observations excluded from every training step.Objectives: This methodological study specifies a sequential Hybrid Hierarchical Machine Learning (HH-ML) framework for behavioral session segmentation and cross-fitted purchase-intent modeling.Methods: The reference dataset is the UCI Online Shoppers Purchasing Intention benchmark, with 12,330 distinct sessions, 17 predictors, and a binary Revenue outcome; 1,908 sessions (15.47%) are purchase-positive. The descriptive segmentation analysis compares standardized engagement-feature K-Means with K-Means on a four-dimensional deep-autoencoder embedding. The proposed predictive design combines a global LightGBM classifier, eligible segment-specific experts, and a logistic second-level model fitted from strictly inner-fold out-of-fold probabilities. The specification defines nested representation fitting, numerical expert-support rules, fixed training budgets, a decision threshold, matched ablations, probabilistic scores, cluster-stability analyses, and paired statistical inference.Results: Previously documented, full-data clustering summaries are separated from any test-fold evidence: reported best silhouettes are 0.494 at three clusters for classical K-Means and 0.658 at two clusters for the embedding approach, a contrast confounded by different cluster counts. The available empirical evidence supports only descriptive session profiles; no comparative purchase-intent performance or calibration result is established by those profiles.Conclusions: This article characterizes a reproducible evaluation framework rather than claiming demonstrated predictive superiority or deployment readiness.