Predicting E-Commerce Purchase Intention Using Machine Learning

Abstract

The fast-evolving digital commerce environment demands precise predictions of consumer buying intentions to develop personalized experiences and boost conversion rates and user satisfaction on e-commerce platforms. The field has extensively utilized traditional statistical models in conjunction with behavioral theories; however, these methods fail to adapt to high-dimensional, imbalanced session-level data. This learning proposes a machine learning-based approach to predict online purchase intention using the widely recognized Online Shoppers Intention dataset. The methodology involves a reproducible pipeline that integrates data preprocessing, the synthetic minority over-sampling technique (SMOTE) to address class imbalance, chi–square–based feature selection, and a comparative evaluation of multiple classification models. The pipeline was tested on a 70:30 train-test split using stratified sampling to maintain class distribution, and further validated through 10-fold cross-validation to enhance robustness. The Support Vector Classifier (SVC) was found to be the best model in terms of both ROC-AUC and F1 score, achieving an ROC-AUC of 0.886 and an F1 Score of 0.633, thereby efficiently discriminating between purchase and non-purchase sessions. We also explore Random Forest, Decision Tree, and Ridge Classifier models to support a more holistic understanding of performance across a variety of complexity and interpretability levels. Importantly, the research also uncovers important behavioral predictors, including product-page engagement and whether the visitor is a returning one, providing interpretable insights that are consistent with real-world e-commerce practices. These results suggest the potential for implementing machine learning models for real-time behavior forecasting in online retail contexts and show that a data-driven pipeline can add value to traditional behavioral modeling counterparts.