Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce (1906.12128v1)

Abstract: With the rapid growth in fashion e-commerce and customer-friendly product return policies, the cost to handle returned products has become a significant challenge. E-tailers incur huge losses in terms of reverse logistics costs, liquidation cost due to damaged returns or fraudulent behavior. Accurate prediction of product returns prior to order placement can be critical for companies. It can facilitate e-tailers to take preemptive measures even before the order is placed, hence reducing overall returns. Furthermore, finding return probability for millions of customers at the cart page in real-time can be difficult. To address this problem we propose a novel approach based on Deep Neural Network. Users' taste & products' latent hidden features were captured using product embeddings based on Bayesian Personalized Ranking (BPR). Another set of embeddings was used which captured users' body shape and size by using skip-gram based model. The deep neural network incorporates these embeddings along with the engineered features to predict return probability. Using this return probability, several live experiments were conducted on one of the major fashion e-commerce platform in order to reduce overall returns.

• The paper presents a dual-model deep neural network that predicts return probabilities at both the cart and product levels.
• It leverages advanced feature engineering with product embeddings from matrix factorization and user characteristics from skip-gram models, achieving an AUC of 83.2%.
• Live experiments demonstrated up to a 4% reduction in return rates through targeted pricing and promotional strategies.

Analyzing Return Prediction in Fashion E-Commerce

The paper "Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce" presents a sophisticated approach to predicting product returns in the fashion e-commerce sector. With the increasing prevalence of liberal return policies in e-commerce, the financial toll of returns is considerable, encompassing reverse logistics and potential product damage. This research introduces a method to predict return probabilities even before an order is placed, allowing for preemptive actions that aim to mitigate these costs.

The core methodology leverages a Deep Neural Network (DNN) to predict return likelihood at both cart and individual product levels. This is achieved through product embeddings derived from Matrix Factorization (MF) using Bayesian Personalized Ranking (BPR) and user characteristics captured by skip-gram models. These embeddings are combined with engineered features to deliver return predictions. The novelty of this paper lies in its dual-model approach, providing predictions at two granularity levels—cart and product—to facilitate a range of targeted actions.

Empirically, the research demonstrates significant predictive accuracy. Utilizing an AUC of 83.2%, the model substantially improves upon baseline gradient boosted classifiers. This high degree of precision and recall is achieved through rigorous feature engineering, which incorporates product-level, cart-level, and user-level attributes. The results underscore the effectiveness of capturing latent sizing vectors and purchase history, with particular emphasis on size and fit, a primary cause of returns.

The paper also explores practical implications through live experiments on a large e-commerce platform. The intervention strategies include dynamic pricing such as personalized shipping charges, coupon incentives for non-returnable items, and innovative purchasing options like 'Try and Buy.' These approaches yielded measurable reductions in return rates—up to 4% in certain scenarios—demonstrating the model's utility in real-world applications. Notably, the integration of this AI-driven insight allowed for refined customer segmentation based on predicted return behavior, which can inform inventory management and customer engagement strategies.

Future developments could extend this predictive framework to broader e-commerce settings and explore additional preemptive actions that leverage real-time user-product interactions. While the proposed model shows considerable promise, further exploration into dynamic online learning models may enhance predictive capabilities, accommodating evolving consumer behavior patterns. Overall, the research presents a substantial contribution to the optimization of e-commerce operations, with implications for improving profitability and customer satisfaction within the industry.