08/17/2023

Graph Learning: Powering the Fashion Experience

Building a Great User Experience in Fashion is Hard

The world of ecommerce today is predominantly powered by machine learning to optimize the user experience and drastically improve how people interact with and consume goods and services. The default expectation from consumers is that their needs will be anticipated and they will find what they need quickly and effortlessly. From personalized home pages, to highly effective search and discovery, to retail and operational optimization, machine learning is powering the online shopping experience.

Fashion is no different – but the dynamic nature and landscape of the fashion world makes this more challenging. Fashion is ever-changing with new trends going in-and-out of style, brands have their own unique messaging and ethos that resonates with particular types of individuals, different geographies and demographics have starkly different needs and preferences, and consumer needs are ephemeral (ie users have a particular need for something based on events or activities going on at a particular time in their lives – that need vanishes as soon as a vendor meets that need or the time has passed). What’s more, style and compatibility across products is often hard to compute and quantify in an ever-changing landscape. But if you want to anticipate a single user’s needs, you need to factor all of these elements into account. To do this for all users, you need a platform that can do this at scale.

Why Fashion Needs Graph Learning

Graph learning, and more specifically Graph Neural Networks (GNNs), is a relatively new subset of ML designed to overcome these challenges.

Your online platform is a heterogeneous graph represented by users, their attributes and activities, products, and the corresponding transactions. When you apply GNNs to this graph, the model captures the relational structure between these entities. With each new prediction, GNNs leverage signal from the entire graph, including information multiple hops away. This provides comprehensive context that a traditional ML model would not have.

GNNs can build a better representation of your users and products. Each customer has a very unique profile that is the result of past behavior (previous searches, interactions with product pages, surveys, recent purchases), learned attributes (preferences, favorite hobbies/activities), and point-in-time occasions which might require specific products by a certain date. Similarly, each product has a unique profile captured by descriptive tags and categories (color, fabric, print, style, brand) which matches specific preferences and occasions. 4

GNNs can interpret this information at scale to provide the best predictions at the right time. So where is this applied in practice?

Home Page Personalization

Graph learning makes personalization at scale more effective. This means a home page that updates in real-time with uniquely curated products for each user. Each product placement is the result of many predictions across categories identifying what a user is likely interested in at that particular point in time.

Context about that user is gathered from various sources to deliver highly relevant products. Even in the case of a new user with little to no historical data on your particular platform, GNNs leverage signal from other parts of the graph to infer meaningful products based on similar users’ preferences.

Product Discovery and Search

When a user goes to the search page, they’re usually looking for something specific. Graph learning makes it possible to surface the most suitable products at the very top based on high relevance scores. A user’s query is combined with context from all other similar users and previous interactions to provide a highly accurate recommendation. Similarly, GNNs can take this context to predict “Suggested searches” for each user – based on inferred preferences, you can prompt the user to search for something specific they would likely purchase. This also makes it possible to surface new products or categories that the user has a high affinity for.

Cross-selling and Compatibility

One of the biggest questions fashion retailers ask is: “which item should we select to match the product a user is looking at?”

The ability to connect users to new items based on product compatibility requires learning multi-hop connections that capture both how different products interact with each other, and how different users interact with similar products. By taking the context of the entire graph, you can deduce co-purchasing habits of specific users and when these habits translate to new users.

Relational patterns can be learned at scale – for every user looking to complete a purchase, GNNs will recommend compatible items to match the occasion, style, and preference.

Retail Strategy

One of the most undesired outcomes for fashion retailers is ending up with unsold inventory – this occurrence is very common. By analyzing user demographics and purchasing behaviors across store locations, a fashion retailer can make forward-looking predictions on which lines will perform well given the store and how much to stock. In 2018, H&M’s lack of understanding of consumer preferences resulted in a loss of $4B in unsold inventory. By leveraging graph learning, they quickly realized their inventory stocking strategy was severely lacking. They were able to optimize their stock for the demographic while cutting down on over 40% inventory.

By predicting both future demand and likelihood to purchase, GNNs can combine these predictions to bridge the gap between expected sales and allocated inventory. This translates to many other forward-looking strategies based on future user behavior at scale, optimizing the business through advanced forecasting.

Where do you get started?

While GNNs are the most powerful tools to optimize a fashion business, they’re not easy to implement. You can read more about the benefits and the challenges of implementing GNNs in production for personalization.

For these reasons, we introduce the Kumo platform. You can read more about our platform and what Kumo has done for ecommerce at large, but we encourage you to reach out directly to learn more and try it yourself!