Meta-Learning for Size and Fit Recommendation in Fashion

Abstract

Fashion e-commerce has enjoyed an exponential growth in the last few years. A key challenge of the market players is to offer customers a personalized experience and to suggest relevant articles. In that respect, although product recommendation is a well-studied field, size and fit recommendation is still in its infancy. The size and fit topic is a very challenging problem as data is extremely sparse and noisy. Most approaches so far have exploited traditional machine learning techniques. In this work, we bring forward a meta-learning approach using an underlying deep neural network. The advantage of such an approach lies in its ability to exploit large scale data, learn across fashion categories, and absorb new data efficiently without retraining. We benchmark our method against 3 recent methods proven successful in the domain, and demonstrate various strengths of the proposed approach. To that end, we use a large-scale anonymized dataset of about 9.4 million customer-size interactions, collected over 5 years from around 384k customers.

SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce

“…With the aim of unifying the different sizes across size systems and brands, [9] propose a method to automate size normalization to a common latent space through the use of articles ordered data, generating mapping of any article size to a common space in which sizes can be better compared. More recently, there has been emerging research addressing the problem of personalized size recommendation for online fashion retailers [1,7,11,13,20,21,[26][27][28]. Given the order history of a customer (or personal customer data such as age, weight, height, etc.…”

“…To that end, we present a Bayesian model which uses size-related return data from customers to learn which articles would fit normally and which would exhibit a size issue. Unlike previous work [1,7,11,13,20,21,[26][27][28], our approach extensively leverages the size-related return rates of articles to greatly focus on modeling articles sizing behaviour. Although the (weakly annotated and subjective) return data from customers is leveraged in the model, the latter is agnostic to the specific customer who places the order and thus by design does not provide a customer with a personalized size recommendation; instead it informs them about article specific sizing characteristics.…”

“…Although the (weakly annotated and subjective) return data from customers is leveraged in the model, the latter is agnostic to the specific customer who places the order and thus by design does not provide a customer with a personalized size recommendation; instead it informs them about article specific sizing characteristics. Thanks to this strategy we create an algorithmically driven customer experience that presents a very low cognitive load by purposefully focusing our approach towards the articles themselves and the underlying manufacturing and brand characteristics, in contrast to personalized size recommendations [1,7,11,13,20,21,[26][27][28] which are emotionally engaging for the customers by trying to convey to them "what their size is"-strongly increasing the chances of a correct recommendation being dismissed altogether [30]-our approach rather focuses on the articles themselves and supports customers to make an informed decision on which size to order given the inferred size and fit characteristics of a given article. This in turn, gives customers the flexibility to adapt their choice based on their fit preference instead of having to cope with a rigid recommendation engine telling them what their size is.…”

“…This in turn, gives customers the flexibility to adapt their choice based on their fit preference instead of having to cope with a rigid recommendation engine telling them what their size is. Therefore, in contrast to [11,20,28] which evaluate the performance of their methods solely based on the accuracy of the personalized size recommendations (i.e. whether the system is good at inferring a customers' size), in this work we go a huge step forward and evaluate the impact of our system on the end-to-end size and fit journey; beyond onsite and unboxing stages and at the size-related return level.…”

“…Considering the above challenges, previous work [11,20,27] has been mainly focused on evaluating the quality of the recommendations not on their impact on reducing size-related returns, but rather on a combination of customer based metrics through A/B tests [34] or continuous evaluation frameworks. Such customer based metrics include monitoring a change in customers' conversion rate, in the number of products added to the cart, in the revenues per visit, in selection orders (where a customer orders the same article in multiple sizes), and in reorders (where a customer returns an article and reorders it in a different size).…”