on’t you hate those moments when you are standing in front of a mirror in a clothing store’s fitting room, trying on a new outfit, with multiple thoughts rushing through your head: Is this still in fashion? Where will I wear this? Can I afford it?
Despair no more. Stitch Fix, a popular online subscription and personal shopping service, promises to spare its customers from the drama of shopping by matching each person with a personal stylist who selects clothing and accessories based on the individual’s size, style and budget. How can a stylist, who does not personally know you, manage to successfully curate your wardrobe?
SEE MORE FROM PART ONE OF A SPECIAL SERIES ON AI
The secret sauce is the algorithms, which are at the core of the company’s business model and do everything from drive the clothing selections to assign human stylists to optimize production and logistics. As a personal style service “that evolves with your tastes, needs and lifestyle,” Stitch Fix benefits from algorithms on a daily, customer-by-customer basis. It’s one thing to fill a clothing order for a client but quite another to do the heavy lifting and decision making for them.
In this article, Eric Colson, Stitch Fix’s chief algorithms officer, takes us on an algorithms tour, explaining how the fickle business of fashion runs like a smooth machine when driven by algorithms.
Data Science: Algorithms 101
What exactly is an algorithm? It’s a set of instructions, with the simplest one being an if-then formula. For example, if I push the number 12 button in the elevator, it will take me to the 12th floor. If I send a customer a blue dress, she will certainly like it. Well, not so fast. Such simple algorithms, called sorting algorithms, are not enough in fashion or business at large.
This is why Stitch Fix uses much more complex and nuanced formulas for its algorithms, with the machines trying to figure out the relative likelihood that a particular client will like a certain piece of clothing. While the company does not publicly disclose all the types of algorithms it uses, some of these formulas are collaborative filtering problems (e.g., those who have liked what you have liked have also liked…). The company also uses mixed-effects modeling algorithms, which incorporate longitudinal data into complex statistical models, allowing Stitch Fix to get a sense of how fashion tastes are evolving over time, both for individual customers and for its customer base as a whole.
Of course, no matter how many formulas are at work, there can never be certainty that the customer will like and buy the algorithm-selected item of clothing. But the more information algorithms have about client preferences, the more likely they are to eliminate uncertainty and make a perfect choice.
Feeding The Algorithms With Data
At the core of a well-functioning algorithm is a proprietary set of data. At Stitch Fix, the first data set comes from the customers themselves, who fill out an in-depth profile when joining.
Questions range from fundamentals (height and weight), to taste and preferences (do you like your shirts tucked or untucked?), to personal traits (are you a risk taker?) and lifestyle (are you a new mom?).
The second set of data is about merchandise. It ranges from measurements (about 30 points of measure for a men’s shirt), styles (bohemian, chic) and material (does it wrinkle easily? does it need to be dry-cleaned?). Each item of clothing is tagged multiple times with match scores from different algorithms derived from client preferences, and then ranked.
Having the data from both sides, a set of machine learning algorithms can match a client to the best-suited merchandise. These algorithms figure out how much each characteristic matters to a particular person, and how to trade off one for the other (e.g., a new mom probably does not want to buy casual clothes made from fabrics that need to be dry-cleaned).
How can a stylist, who does not personally know you, manage to successfully curate your wardrobe?
As time goes by, algorithms learn from actual customers, both individual and in aggregate, how to think about clothes. This is possible thanks to the feedback data collected from customers, which is transmitted back to the algorithms so that they can see how their decisions worked in real life—and use this information to constantly improve their decision-making formulas (machine learning).
“Feedback data is the most valuable data set for us, our secret sauce, and unique to us,” says Colson. Traditional retailers may never know why a customer walked out of the store without buying anything, and thus don’t know what to fix to make the customer happier next time. Stitch Fix asks its customers many questions about their interactions with clothes, trying to understand the reasons for buying or not buying.
“The copacetic relationship we have with our clients only works if they get value from us. There’s no ‘selling’ here—only relevancy,” says Colson.
Of course, it’s not always obvious which questions to ask to discover the most critical factors in buying clothes. Yet Stitch Fix captures such material information with text boxes in which customers can record their comments. Natural language processing of these comments can then point to new clusters of characteristics that become part of algorithms. For example, a comment from a client read, “I was happy because I could wear this garment to the park and then to an outdoor wedding,” which points to the importance of having casual clothes that can be worn to formal occasions. Colson points out that a vast majority of Stitch Fix algorithm capabilities, which are significant for business outcomes, are conceived by the data science team.
Algorithms Running The Business Of Business
No matter how good the customer experience, it won’t do a company any good if it is not delivered profitably. “Algorithms bring efficiencies that make us more profitable, and we have implemented them pervasively throughout the company,” says Colson. Here are a few other ways algorithms are driving Stitch Fix’s business forward.
“The copacetic relationship we have with our clients only works if they get value from us. There’s no ‘selling’ here—only relevancy.”
1. Matching the right stylist with the right customer
Ultimately, at Stitch Fix, it is a human stylist who finalizes the clothing selections and even writes a personal note describing how the client might accessorize the items for a particular occasion and how to pair them with other clothing in his or her closet.
Humans are more heterogeneous than machines. The machines are all the same. But human stylists are going to be better suited to some clients than to others. To pair the right stylists with the right customers, Stitch Fix calculates a match score between each available stylist and each client who’s requested a shipment. This match score is a complex function of the history between that client and stylist (if any), and the affinities between the client’s stated and latent style preferences and those of the stylist. By playing matchmaker, Stitch Fix connects clients with the styles—and the stylists—they love.
2. Better inventory control
Inventory management is one of the trickiest parts of the retail business. How many different styles to order, in what sizes? Traditional retailers usually make educated guesses, or buy in some bell-shaped distribution curve of sizes. “We look at our clients, figure out how many of them are going to be interested in this particular dress, and then we buy in the size distributions that they exhibit,” says Colson. “This makes inventory much more manageable.”
3. Lower transportation costs
The costs of transportation can also eat deeply into many retailers’ profits, as items travel to different pickup and delivery locations across the country. At Stitch Fix, algorithms calculate a cost function for each warehouse based on a combination of its location relative to the client and how well the inventories in the various warehouses match the client’s needs, thus cutting down on unnecessary transportation costs.
4. Designing new styles
The fashion industry has been built around the persona of a creative and often fickle designer, whose inspiration gives us new styles every year. Just seeing the designer’s name on a tag can command hundreds of dollars. But do we always feel that these new styles are designed for us?
At Stitch Fix, algorithms help design new styles that are tailored for particular client segments that tend to be underserved by other brands. Designer algorithms use the same mechanisms to develop new styles as mother nature does in evolution by natural selection: recombining attributes from existing styles and possibly mutating them slightly, then testing for “fitness” (customer feedback on similar styles).
Well, almost the same. If it was operating exactly like nature, random recombinations and mutations would be thrown out there to live or die. Stitch Fix wants as many of its creations as possible to live. That’s why the company takes steps that mother nature doesn’t and is somewhat pickier about what makes it into its inventory. After highlighting a variety of attributes that have a high probability of love, Stitch Fix works with its human designers to vet and refine this collection, and ultimately produce the next generation of styles.
So don’t be surprised if the next time you pick up a clothing item, the tag says that the designer is: The Algorithm.
Stitch Fix is a prime example of how companies can find unique ways to leverage algorithms. And by building its entire business model around algorithms, it has made itself an attractive place for data scientists to work. Of course, all this requires an investment of time, money and brainpower—but the payoff is worth it. Companies that neglect this investment, or delay too long, may one day watch their business model unravel.