An Explainable Personalization Framework for E-Commerce Based on Cart Purchase Prediction and Item-to-Item Recommendation with External Validation

An Explainable Personalization Framework for E-Commerce Based on Cart Purchase Prediction and Item-to-Item Recommendation with External Validation

Parisa GhahremanShahraki1

1) Computer Engineer

IOI Code : AAHC-ABKKAD
Publication : National Conference on Management, Commercialization, Knowledge-Based Economy and Investment for Production(impeb.ir)
Abstract :
    Effective personalization in e-commerce requires the integration of purchase behavior prediction, relevant recommendations, and the explainability of algorithmic decisions. In this study, an explainable framework for e-commerce service personalization was developed by integrating three main components: prediction of the purchase probability of an item added to the cart, behavioral item-to-item recommendation, and a probability-based action policy. The primary data comprised 2,755,641 cleaned events from the Retailrocket dataset. The learning task was defined as predicting the purchase of the same item within 14 days after its first add-to-cart event. After excluding carts without a complete observation window, 56,385 samples and 23 leakage-free features were constructed at the user, item, and user–item interaction levels. Evaluation was performed using a temporal split, and Logistic Regression, Decision Tree, Random Forest, CatBoost, and XGBoost models were compared. CatBoost was selected as the primary model and achieved a ROC-AUC of 0.7584 and a PR-AUC of 0.5947 on the temporal test set. Using a threshold of 0.25 selected on the validation data, recall and F1-score reached 0.7228 and 0.5575, respectively. SHAP analysis showed that user transaction history, the user’s prior purchase share, and prior interaction patterns with the same item were among the most influential predictors. For external validation, an equivalent framework was evaluated on the session-based YooChoose dataset, obtaining a ROC-AUC of 0.6337. In addition, the item-to-item co-occurrence recommender, using a hybrid popularity-based fallback strategy, achieved a Hit Rate@10 of 15.26%, substantially outperforming the popularity baseline. The findings indicate that combining purchase-probability prediction, behavioral recommendations, and explainability can provide a practical decision-support prototype for low-cost and interpretable e-commerce personalization.

Keywords : e commerce recommender system personalization explainable artificial intelligence catboost shap cart purchase prediction