Attention Gets You the Right Size and Fit in Fashion

Abstract

Avoiding returns in e-commerce platforms has become a critical issue in terms of both increasing customer satisfaction and decreasing carbon footprint. In the online fashion industry a very large part of the returns is due to size and fit issues that arise from the underlying complexities of shoe and garment manufacturing combined with subjective preferences of customers towards what fits them best. In this context, size recommendation systems capable of estimating a customer’s size in thousands of available brands and categories ahead of purchase time are deemed invaluable in dramatically reducing the number of returns related to size and fit. We present a flexible and scalable size recommendation approach that overcomes some limitations of current state-of-the-art work by building upon recent advances in natural language processing and casting the size recommendation problem as a kind of “translation” problem (from articles to sizes) using an attention-based deep learning model for size and fit prediction. Through extensive experimental results, over millions of customers and articles, we demonstrate how this approach is capable of dealing with multiple customers buying from a single account, leveraging cross-category and temporal information to make better predictions, and providing explanations on the final size predictions it produces, thereby helping reduce the potential emotional costs of such predictions for customers.

Alex Zhao—Work done while an intern at Zalando SE.