Last summer, we wrote about Stitch Fix’s early experiments in data driven fashion design. Since then, we’ve been studying, developing, and testing new ways to create clothes that delight our clients. Some of this work was featured yesterday in an article in The Wall Street Journal. As a companion to that piece, we wanted to highlight a few avenues that we have explored recently.
In our last post, we described an iterative design approach inspired by biological evolution; we conceived of a population of blouses in our inventory. From this population, we selected sets of high fitness individuals (i.e., blouses our clients loved) and we recombined them (i.e., mixed and matched their best attributes). This produced a new generation of children blouses, whose attributes were occasionally mutated to rare attributes in our inventory (i.e., switching to a less common hemline, pocket type, etc).
Since our last blog post, we have been studying new types of mutations. For example, can statistical modeling identify when a successful blouse has an attribute that is holding it back? If so, can we suggest a mutation that replaces the underperforming attribute? To illustrate, can we identify when a parent blouse is successful despite its leopard print, and then change it to the floral print that everyone loves this season?
We are also examining how we can leverage less structured types of data. For example, can we extract features from images of blouses or the text feedback that clients provide in response to a blouse? If so, can we use these features to drive recombination or mutation recommendations? To illustrate, can we extract nuanced labels about color palettes (e.g., warm vs. cool vs. deep vs. pastel) from images of blouses, and then learn about the tones that different clients prefer at different times of the year?
Finally, we are testing and learning about the unique challenges that emerge as we design in categories other than blouses. This provides an intriguing opportunity to assess the generality of our data, models, processes, and even our fundamental approach. For example, is fine grained data about fabric type (e.g., yarn gauge) more critical when designing sweaters; is the relationship between color and client feedback moderated by product category (blouse vs. skirt); is it critical to target a client’s body shape or fit preference when designing a dress? Answering these and many additional questions will help us unlock the power of data-driven design.