Data Innovators Eric Colson

Published on May 9th, 2016 | by Joshua New

0

5 Q’s for Eric Colson, Chief Algorithms Officer at Stitch Fix

The Center for Data Innovation spoke with Eric Colson, chief algorithms officer at Stitch Fix, a subscription shopping company based in San Francisco. Colson discussed the challenges involved in teaching an algorithm to understand something as subjective as style, as well as how the company’s data on customer habits allows it to find clothes that customers enjoy even if their preferences might indicate otherwise.

This interview has been lightly edited.

Joshua New: One of Stitch Fix’s selling points is that the clothes customers receive as part of the subscriptions services are selected based on algorithmically generated recommendations. What goes into quantifying something subjective like style so an algorithm can make these recommendations?

Eric Colson: Good question. Clothes are an extremely personal thing for consumers, so it’s important that you both cater to their individual preferences but also factor in social norms, such as current fashion trends. You have to balance in that bit of conformity with people’s need for individual expression. Fortunately, through our business model we have this nice symbiotic relationship with our customers. Because they’re coming to us for us to pick out clothes for them, they understand that they’re going to have to tell us something about themselves. Before their first shipment, they fill out what we call a a style profile—50 questions that tell us things about yourself, your fit and style preferences, and even your penchant for taking risks, as well as standard things such as height, weight, and size. We ask them their preferences upfront, sometimes by showing them images of clothing and asking them questions like, “Is this you?,” or “Is this something you would never wear?” Sometimes it’s hard to get a customer to express what their style is, but these images usually get better responses. We also allow them to create albums from Pinterest that demonstrate what their style is without having them articulate it.

So we get this data from customers upfront very explicitly, which helps us in the beginning by giving us a great head start on picking out clothes we think they would like. Each shipment we have is called a “Fix.” On the very first Fix, we start getting feedback as customers can try these clothes on in the privacy of their own home and ask their friends and loved ones about different pieces. This is really rich feedback, since they can evaluate clothes based on how they actually fit and look on their body instead of just making a simple decision about what they see on a page in a magazine. Customers will give us some structured information, such as feedback about the fit, the price, the style, and the size, as well as unstructured information—we have a text box they can fill out with any additional thoughts, like “this fits great on the body portion, but it’s a little tight on the shoulders.” This rich detail is incredibly helpful as our algorithms learn a customer’s preferences.

The nice thing is that we can aggregate our customers’ preferences together to get even more precise. For example, if one customer says a certain piece runs a little bit big, we can compare that to other responses to see if others say the same thing so we can adjust our own data on that piece. So we benefit from both individual responses and more crowd-sourced feedback.

New: ”Chief algorithms officer” is a pretty unique title. Why the focus on algorithms specifically, rather than something more standard, such as “chief data officer” or “chief data scientist?”

Colson: When I joined the company, I had to figure out which title would be appropriate. Your title connotes the type of work you do and, importantly, influences who you’re going to be able to recruit. For instance, on LinkedIn, your title can tell someone what you’re all about. So the obvious candidates were things like chief analytics officer, which many companies have, but analytics connotes more business intelligence-type tasks than what we’re really doing. Chief data officer often evokes more of an IT function of managing data, but not necessarily using it, so that wasn’t right either. Algorithms just seemed right, because it’s not so much just having the data and getting insights from data, but putting insights into action.

Chief data scientist would have worked, but I was always a little bit worried about that phrase. I think it’s great and makes a lot of sense to me, because for years I struggled with the “what am I” question—I’m not really an engineer, I’m not really a statistician. So when the phrase was popularized by Jeff Hammerbacher and DJ Patil, I thought, “yes, finally someone figured out how to describe what I do!” When I worked for Netflix, I named my team “data science and engineering,” but with some trepidation. I was concerned that this phrase gets played out and used to describe things that data scientists don’t actually do. I shy away from it because I think it gets abused in industry to the point where you no longer know what it means.

So chief algorithms officer was the right fit. We have algorithms all over the company, for style recommendations but also to manage inventory, to support customer outreach, and dozens and dozens of other places. We now have a team of over 80 algorithm developers.

New: You mentioned that Stitch Fix can fine-tune its recommendations if a customer gives you access to their social media pages like Pinterest. How does this work?

Colson: This is part of our style profile, so when you sign up, there’s a box you can check to share your social media with us—your Pinterest and Instagram pages. Most customers do, and as I mentioned this is a really great way for customers to express things they can’t really put into words. They can see stuff online, think “that’s great,” and share that with their Stitch Fix stylist. They often make a specific page on Pinterest, called a Pinboard, specifically for Stitch Fix to tag all their apparel-related activity.

The interesting thing is that we have both human stylists, which really digest the hard-to-grasp aspects our customers are trying to convey to us in these images, and also an experimental effort to use computers to analyze these images. Our algorithms will analyze an image and compare it to our inventory to find similar items to recommend to our stylists, which is great because human stylists are great at perception, but not as great at searching through tens of thousands of items in our inventory to find a good match.

New: Style, as you mentioned, is very subjective, so even if you amass enough data to develop a specific style profile for a customer, there’s still a chance he or she won’t like an article of clothing despite all of your data indicating otherwise. How do you train your algorithm to accommodate the fact that customers may be picky for unquantifiable reasons?

Colson: This is a challenge, but it’s expected. We pick clothes based on stated preferences that we think are likely to meet your needs. We send them to you, you try them on, and then we get your feedback, which is incredibly valuable. From there, regardless of if the feedback is positive or negative, that helps us. It’s just like in real life if you were to walk into a store and the stylist suggests a pair of leather pants you don’t like. If you say “no, that’s not for me,” we can then eliminate wide swathes of inventory that won’t be relevant to you. We may not always get it perfectly right, but this feedback helps us with individual clients as well our other clients with similar tastes so we can continuously improve.

New: Stitch Fix launched in 2011. What have you learned about how people shop for clothes based on the data you’ve accumulated so far?

Colson: I think one of our most interesting findings is that people just can’t seem to judge clothing from images alone. You can recommend something on a web page, but the customer might just say “no, that’s not for me.” But if you send it to them and say “trust me, try it on,” which we can do since we have the data to back it up, often times you’ll find that customers are surprised at how much they like it. That bit of knowledge, which we learned in 2012, was astounding to us. So many things that would have been rejected if a person were to just judge from an image, or even walking by it in person, we’ve found customers really enjoy if they try it on. We call it “surprise and delight,” and it’s been a game-changer for us.

I witnessed this first hand back when I was an advisor for Stitch Fix before I started working here. I asked my wife if she would try the company out and explained how it would send her clothes based on her preferences. She got her first Fix and the first item was a scarf, to which she said “ugh, I don’t want a scarf.” But then she touched it and tried it on and looked at herself wearing it in the mirror, and was so surprised at how much she liked it. By having it there right in front of her, her opinion about it completely changed. So our data has shown us that this process, called an experiential trial, pushes customers a bit out of their comfort zone and can be incredibly effective. Our customers come to us for the convenience, and they stay for this surprise and delight of finding clothes that they would never have thought they would have liked. It’s obvious to us now, but back in 2012 that was a huge surprise and proved that people really do need help with this.

Another interesting thing we learned is just how unique everybody is. As retailers, we love to develop things like personas—a fictitious character with certain traits that represents a group of customers. So we looked at the data and tried to figure out what the most representative persona of our customers would be. We had this idea in our head that this persona would fit some particular demographic. But when we looked at the data, we found that there really wasn’t any one segment that represented even one percent of our customers. Everyone really is just that different. We had sent millions of Fixes by this time, so we looked at the shipment data and found that every single Fix was unique. At the time, we only sold women’s clothes, and not one Fix had the same five items that another woman got. This was stunning to us. We make no attempt to mandate that all orders are unique—we simply cater to preference, and each customer has particular preferences. It makes sense to us now, and we have had customers receive the same order a couple of times since, but it really made us realize just how unique every customer is and it was something we never would have guessed before we started.

Tags: , , , ,


About the Author

Joshua New is a policy analyst at the Center for Data Innovation. He has a background in government affairs, policy, and communication. Prior to joining the Center for Data Innovation, Joshua graduated from American University with degrees in C.L.E.G. (Communication, Legal Institutions, Economics, and Government) and Public Communication. His research focuses on methods of promoting innovative and emerging technologies as a means of improving the economy and quality of life. Follow Joshua on Twitter @Josh_A_New.



Back to Top ↑

Show Buttons
Hide Buttons