From day one, we’ve been obsessed with making it easy for you to find what you love. Fast forward to today, and our unique combination of art and data science now powers a whole ecosystem of experiences that deliver the magic of personal styling at scale to more than 4 million clients across the US and UK – whether it’s the Fix curated by your stylist, the shoppable community inspiration on your personal feed or the curated assortment of items in your Freestyle shop.
Understanding personal style at this kind of scale is incredibly complex. Not only is style highly nuanced and personal, but it is incredibly dynamic: trends come and go, and our individual style can evolve and change over time.
To gain a holistic understanding of style – of both our clients and our merchandise – our teams developed an algorithm that we call “Latent Style” using billions of ratings from millions of clients in Style Shuffle and the insights from their Style Profiles. This is one of our foundational algorithms: Latent Style can predict what items you will rate positively in the future, which informs everything from what items we recommend your stylist puts in your Fix or in your personal shop in Freestyle, to what items our buyers purchase for our inventory and even the items that are included in emails we send you.
How Latent Style works
When the Latent Style model is trained, it creates a higher-dimensional space, a bit like a map, with a position for every item we’ve shown in Style Shuffle, and for every client – distilling 10B data points into 10M coordinates on the map. We’ve taken these coordinates, further processed them down to three dimensions, and created an interactive visualization for our teams across the business to use to explore the style of our items and style preferences of our clients.
Latent Style enables us to personalize your experience at Stitch Fix because it helps us understand who likes what. And because of the sheer volume of ratings and data points across our ecosystem, Latent Style allows us to make personalized recommendations for you even if you haven’t played Style Shuffle yet. So if you like straight leg jeans and fitted tops, and another client who likes straight leg jeans and fitted tops also likes midi skirts – Latent Style will tell us that you might also be interested in trying a midi skirt.
Latent Style also helps to power our business behind the scenes, allowing us to measure how similar items in our inventory are, the popularity of those items, and nuanced understanding of the style of those items. This enables us to quickly and easily cluster related items in this Latent Style space without manually feeding in traditional merchandising attributes (like department, item type, color, cut, etc.) – which unlocks even more personalized recommendations and helps us build a dynamic inventory that serves the individual needs of our clients’ ever-changing styles.
This interactive Latent Style model is constantly updating as you engage and share feedback with us, and as we continue to add new items for you to rate. With more than 4.5M new ratings from clients in Style Shuffle each day, we’re able to learn about new style trends in real-time and evolve our understanding of style as yours evolves.
Nathan Selikoff, a UI Engineer, shared more about Latent Style and the ways that art and science come together at Stitch Fix at SxSW this year. You can watch part of his talk here.
For more behind-the-scenes about innovative work happening at Stitch Fix, visit our Multithreaded blog.