One Size Does Not Fit All: Predicting Product Returns in E-Commerce Platforms

Abstract

Providing easy and hassle-free product returns have become a norm for e-commerce companies. However, this flexibility on the part of the customer causes the respective e-commerce companies to incur heavy losses because of the delivery logistics involved and the eventual lower resale value of the product returned. In this paper, we consider data from one of the leading Indian e-commerce companies and investigate the problem of product returns across different lifestyle verticals. One of the striking observations from our measurements is that most of the returns take place for apparels/garments and the major reason for the return as cited by the customers is the “size/fit” issue. Here we develop, based on past purchase/return data, a model that given a user, a brand and a size of the product can predict whether the user is going to eventually return the product. The methodological novelty of our model is that it combines concepts from network science and machine learning to make the predictions. Across three different major verticals of various sizes, we obtain overall F-score improvements between 10%–25% over a naïve baseline where the clusters are obtained using simple random walk with restarts.