Personalized Size Recommendation

Research Project by Abdul-Saboor Sheikh & Urs Bergmann

Have you ever wondered what would be the best fitting size of a clothing item that you would like to order online? You may have an idea about the size you generally wear, but this piece of clothing perhaps comes in a different size system or maybe it is a style or category that you have not worn before. Then apart from your own preferences, there may be other factors (such as brand, design, material, manufacturing, etc.) that may as well come into play when determining the right size that would give you the best fit and feel.

Zalando has a lot of customer order and purchase data that spans over a huge assortment of fashion articles. In this project our aim is to benefit from this wealth of information to develop a personalized size recommendation service for our customers.

In spite of having a lot of data, a key challenge with personalization is sparsity – an individual customer often only has a handful of articles in their purchase history. To alleviate this, we are focused on developing models that can leverage size and fit relevant information across individual customers to improve personalized recommendations. Our system is context-aware, where a context may be defined based on attributes that are pertinent to a customer-article pair. The core idea behind the system is to develop a unified mapping scheme such that it allows for both customers and articles to be implicitly projected and compared in a semantic-free (latent) feature space. As illustrated in the schematic below, given such a mapping scheme a customer projected into the hidden space can be aligned with size-specific mappings of an article for making personalized size recommendations.

image1

What machine learning methods do we use? In this project we employ Bayesian theory as well as deep learning; recent advances in deep learning has achieved state-of-the-art in numerous domains. In scenarios with a lot of data and relatively low signal to noise ratio, deep learning has been shown to be empirically quite effective in approximating highly non-linear (unknown) functions. Bayesian theory on the other hand comes in handy when one has to deal with uncertainty that emerges from noise and sparsity in data. It provides us with a nice probabilistic framework for combining beliefs about missing or sparsely observed information together with evidence available in the form of observed data. If you are further interested, checkout the Bayesian size recommendation model that was developed together with our colleagues from sizing team. The work was recently accepted for publication at RecSys 2018 [1].

[1] A Hierarchical Bayesian Model for Size Recommendation in Fashion. Romain Guigourès, Yuen King Ho, Evgenyi Koryagin, Abdul-Saboor Sheikh, Urs Bergmann, Reza Shirvany. RecSys 2018 (to appear).