Towards Waste Reduction in E-Commerce: A Comparative Analysis of Machine Learning Algorithms and Optimisation Techniques for Garment Returns Prediction with Feature Importance Evaluation
Link zur Publikation:
Abstract:
Product returns present a growing economical and ecological challenge for manufacturers and online retailers worldwide. This issue is particularly pronounced in the textile and fashion sector, where over half of all items ordered are returned. One approach to derive measures for mitigating returns and effectively handling inevitable returns is by accurate predictions of returns at the time of order, which allows for improved inventory risk assessment and strategic planning to resell items and prevent garment destruction. This research first investigates the effectiveness of five different machine learning algorithms paired with regularised target encoding for categorical features in predicting returns of a German online retailer, specialised in festive dresses and garments. A balanced accuracy of 0.86 can be reached under the premise of a known customer return behavior. Further, importances of engineered features are assessed using an extended dataset which allows for the comparison of several sales channels. The goal of this research is to develop an AI-driven recommendation system to identify the most sustainable and cost-effective strategies for processing returns for waste reduction and decreased financial strain on retailers.