Data-Driven Analysis of Customer Return Behavior and Fraudulent Patterns in Indian E-Commerce

Abstract

The rapid growth of the Indian e-commerce sector has brought a parallel rise in "wardrobing" and return abuse, where customers purchase items for temporary use and return them for a full refund. Traditional fraud detection systems, particularly tabular models like XGBoost, frequently fail to catch this behavior because they treat every transaction as an isolated event, completely missing the structural "guilt-by-association" patterns inherent in organized fraud. To address this gap, this paper presents a Graph Neural Network (GNN) framework integrated with geospatial hotspot analysis to detect behavioral fraud across a unified Indian market dataset. We benchmarked three distinct architectures to handle this complexity: XGBoost, GraphSAGE, and Graph Attention Networks (GAT). Our comparative analysis revealed that while the XGBoost baseline achieved high overall accuracy, it suffered from a severe bias toward the majority class, missing 61% of actual fraud cases. GraphSAGE emerged as the champion model; by utilizing inductive learning and neighborhood aggregation, it successfully mapped user-item relationships and improved the fraud detection recall rate by 21% over the baseline. Conversely, the more complex GAT architecture struggled with the dataset's extreme class imbalance and overfit the noise. Ultimately, this research demonstrates that transitioning from flat data to Geometric Deep Learning, combined with regional hotspot tracking, is essential for securing modern retail platforms against subtle behavioral abuse.