Return fraud has grown to be a significant issue in the rapidly developing e-commerce industry. As online shopping platforms continue to expand, several companies have adopted generous return policies in an attempt to boost customer happiness and loyalty. However, this has inadvertently created an opportunity for fraud, in which unscrupulous people exploit return policies to acquire illegal goods or money. Returning damaged or used goods, claiming returns for stolen or nonexistent items, or frequently taking advantage of generous return policies are typical instances. In addition to impacting business profitability, these dishonest tactics erode trust in the e-commerce industry. We offer a machine learning-based methodology for identifying return fraud by examining transaction patterns and consumer behavior in order to solve this issue. Specifically, we focus on Return-Frequency (the frequency with which a customer initiates returns), Shipping-Return-Location-Match (whether the return location matches the original shipping address), and Customer-Loyalty (a measure of how long and how reliably the customer has interacted with the platform). Key behavioral features gathered from return histories are used in our technique. Several supervised learning models that estimate the probability of return fraud are trained using these features. Five popular classification techniques are used and assessed: Random Forest, Decision Tree, Support Vector Machine (SVM), Gradient Boosting, and Logistic Regression. The performance of these models in low-data and moderate-data settings is evaluated using a custom dataset of three different sizes (50, 100, and 500 records). Accuracy, precision, recall, F1 score, and AUC ROC are evaluation measures that, when combined, offer a thorough view of model reliability, discrimination ability, and the ratio of false positives to false negatives. The experimental results show several notable trends. Because it can handle smaller data volumes and maximize judgment boundaries, the SVM approach performs best on small datasets (50 records). Tree-based models, such as Decision Tree, Random Forest, and Gradient Boosting, perform better as dataset sizes grow, demonstrating their capacity to represent intricate, non-linear relationships in customer behavior data. To greatly increase the accuracy of fraud detection, we employ a Voting Classifier to build an ensemble learning model that combines the prediction powers of all five fundamental models. By combining the advantages of each algorithm and addressing the shortcomings of each model separately, this ensemble technique produces output that is more consistent and balanced. Notably, the ensemble model earns the best AUC ROC and F1 Score—important measures of classification confidence and overall performance—and performs better than any individual method across the majority of metrics, particularly on the bigger dataset (500 records). Even while separate models have different advantages depending on the size and complexity of the dataset, our research shows that ensemble learning provides a dependable and scalable approach for detecting fraud in real-world e-commerce scenarios. The model’s increased prediction capabilities and adaptability to different data scenarios make it a valuable tool for incorporation into fraud detection systems. Future research can focus on applying this framework to larger, real-world datasets, adding more contextual features like user device information, time-based activity patterns, and natural language return reasons, and exploring deep learning techniques in order to improve automation and accuracy.