Our Recommender

The Mapp Fashion Recommender is a collection of Machine Learning algorithms that process various sources and types of data to produce garment and outfit recommendations.

Fashion Domain

The fashion domain presents unique challenges and has characteristics that are very different from traditional recommendation domains like movies and music. Some of the unique challenges are:

Item catalog changes rapidly: things learned from out-of-stock products are only useful if they can be related to new products
Timing is key: trends and other factors have a big influence. What the customer wants to buy today might be very different from how they have been behaving over the last 12 months. We need the ability to recognize changes in user preference and respond quickly. Compatibility of purchases is crucial: what a customer owns and what she wants to wear has a major influence on the purchase decision.

Data

To address these challenges, the Mapp Fashion Recommender uses data that can be categorized into three groups.

User-Garment interactions: all the various ways in which a user can interact with a garment on the website, in a mobile app, or in an email.

• Purchases, Add to Basket, Wishlist

• Views, Bookmarks, Thumbnail Impressions

• Likes & Dislikes, View Outfits

Explicit User Features: the answers that the user has given to questions in the Mapp Fashion Signup process.

• Shape: height, weight, body proportions, etc.

• Body: colors, reveal and conceal preferences, etc.

• Lifestyle & Attitude: fashion confidence, adventurous, favorite brands, etc.

• Feature preferences: "never wear high heels", "never wear maxi dresses", etc

Explicit Garment Features

• Up to 35-50 features per garment, depending on category and complexity

• Includes things like Fit, Style, Colour, Pattern, Brand, Neckline, Length, Sleeve Detail, etc.

• Labeled by a combination of human stylists and machine learning

Algorithms

The recommender is a combination of several algorithms that learn from this data in different ways. All of the algorithms were developed, optimized, and tested specifically for fashion data. The rich garment feature data that we have allows our algorithms to learn what garment characteristics are preferred by which user and leverage this information to make better predictions. We are able to learn from data collected on items that are no longer available and apply the learnings to new items that have just come out. From the user side, knowing the user features allows us to analyze the demographic of each garment to figure out what kind of users might be interested in it. Putting these two together, we are able to look at what kind of users like what kind of garments in terms of explicit descriptive features.

In terms of ML techniques, we use several building blocks that each address different aspects of the recommendation & prediction problem. We use mainly supervised learning approaches supported by some unsupervised techniques. Our algorithms include the family of Collaborative Filtering algorithms (classic CF, Matrix Factorisation, proprietary), Content-based approaches (proprietary, using our garment features and user features), Graph-based techniques (proprietary), Session-based approaches (optimized for real-time speed). All the algorithms we use are adapted and optimised specifically for the fashion domain, we have no components that we use out of the box and very few that we didn't have to adapt heavily. Several of our algorithms are completely novel and proprietary to Mapp Fashion . We have tight bonds to academia and are active in the Recommender Systems research community, which allows us to stay up to date with the newest developments even before they become widely available.

The recommender is constantly evolving; new algorithms and functionality are added as we complete our research & development cycles. We have tried and learned that standard recommendation approaches don't work very well for fashion, and research-level innovation is required to create high-quality recommendations. Every component of our recommender is thoroughly evaluated in offline experiments as well as live A/B tests. All of our algorithm design decisions are verified by running A/B tests that enable us to measure if they give a real uplift or not.