Social Re-Identification Assisted RTO Detection for E-Commerce

Hitkul Jangra, Abinaya K, Soham Saha, Satyajit Banerjee, Muthusamy Chelliah, Ponnurangam Kumaraguru

Abstract: E-commerce features like easy cancellations, returns, and refunds can be exploited by bad actors or uninformed customers, leading to revenue loss for organization. One such problem faced by e-commerce platforms is Return To Origin (RTO), where the user cancels an order while it is in transit for delivery. In such a scenario platform faces logistics and opportunity costs. Traditionally, models trained on historical trends are used to predict the propensity of an order becoming RTO. Sociology literature has highlighted clear correlations between socio-economic indicators and users’ tendency to exploit systems to gain financial advantage. Social media profiles have information about location, education, and profession which have been shown to be an estimator of socio-economic condition. We believe combining social media data with e-commerce information can lead to improvements in a variety of tasks like RTO, recommendation, fraud detection, and credit modeling. In our proposed system, we find the public social profile of an e-commerce user and extract socio-economic features. Internal data fused with extracted social features are used to train a RTO order detection model. Our system demonstrates a performance improvement in RTO detection of 3.1% and 19.9% on precision and recall, respectively. Our system directly impacts the bottom line revenue and shows the applicability of social re-identification in e-commerce.