TL;DR
- Filter first: shrink the product universe using onboarding answers.
- Learn continuously: swipes and OOTDs update preferences implicitly.
- Explainability: every pick comes with a reason tied to your wardrobe.
- Fast and practical to run at scale.
One of the hardest parts of building MODE was designing the recommendation model that powers our Highlights feed. Building a recommendation model for yourself is easy (go save searches on Grailed); building one that can scale to millions of people is an order of magnitude more difficult. Additionally, we wanted our model to have the following characteristics:
High interpretability
We need to be able to explain why the model recommends someone an item.
Speed
Recommendations need to come fast (we hate waiting).
Cost
Not light our runway on fire.
Adaptable
The model needs to learn and adapt to individual user preferences.
After many prototypes (we’ve tried everything from “hot or not” to Elo and good old‑fashioned machine learning), we settled on a combination of filtration and preference learning. But what does any of that mean?
Step 1: Filter the universe
The first step to good recommendations is limiting the universe of products we recommend you items from. Our database is already massive (growing ~20% month over month), and trying to pick out individual items from there is computationally expensive. We avoid this by first filtering our database using what we learn from the onboarding quiz users fill out when they first download MODE.
The goal of the quiz is to learn what you hate in clothing (and never recommend anything with dealbreaker traits), how you want to dress (pre‑seed quality recommendations from the start), and how you shop for clothing. We sort users into three archetypes:
Infrequent, high‑ticket purchases; highly selective taste.
Higher purchase frequency; embraces new styles easily.
Less frequent purchases; focus on quality.
Step 2: Learn your preferences
When combined together, this creates an almost endless feed of appealing items even among users with wildly different preferences. As you swipe on recommendations and upload OOTDs, our algorithm can learn changes in your preferences implicitly and use that information to deliver even better recommendations.
We think that the best form of search is not having to search at all. Discovery on MODE isn’t just different – it’s simply better.