The days of walking miles up and down shopping streets to find an appropriate dress are long gone. Online services like Stitch Fix leverage data and algorithms to help customers find the clothes they love. Combining AI with the experience of expert stylists they can predict the style of shoppers and send them personalized clothing.
June Andrews, Data Science Manager of Style Discovery at Stitch Fix, talked about these algorithms at our previous DSSVirtual and guided us through the process of how Stitch Fix uses data and recommender systems to pick customized clothes for its customers. Learn more about how this works in this post.
Collecting user data for item recommendations
In the very first step of the shopping experience, Stitch Fix collects 90 data points from a user to make an informed decision on what clothes should be shipped to the customer. This is done by letting the users fill out a survey including questions about style preferences, dress size, height, and location. This data is fed into a series of algorithms, which ranks the clothes, shoes, and accessories in an order that the system believes the customer will like most.
The stylist reviews the customer's profile and preferences shared in the survey and picks five items accordingly. This delivery, also called Fix, is then routed to one of Stitch Fix's warehouses where it is picked, packed, and shipped to the customer's home.
In the first Fix, the stylists usually have limited information about the profile, but as the customers give feedback, better recommendations are made for future orders. When customers receive something they do not like, for example, if a fit is too tight or too short or edgy, the information will be helpful to all other customers with similar preferences. The feedback helps to provide smarter recommendations over time as algorithms learn about the customers and the merchandise.
Discovering new styles with the style explorer
Talking about exploring the different tastes of styles, the major challenge for Stitch Fix is to create a style that would do good before it gets into the ecosystem. As a solution, the company created a recommender algorithm called the Style Explorer. The main focus of this algorithm is to keep it minimal, and to train on millions of data rather than billions. This approach works thanks to a feature called Style Shuffle in their app. The Style Shuffle lets users like and dislike images of items and outfits which therefore enables the collection of feedback prior to the shipping of an actual Fix.
The best practice for Stitch Fix is to set narrow but flexible system parameters and not to over-specify a process. Since target metrics are changing all the time, Stitch Fix developed a system that can handle the complexity.
Since Stitch Fix has rich, meaningful, and highly actionable data (client data, merch data, feedback data), it can measure the business context to allow for specific evaluation of performance for different inventory strategies. One more consistent feature in Stitch Fix’s predictive algorithm is - whenever possible, they compute the predictions under all parameter values and ask users to specify the metrics.
Data science and machine learning are embraced at Stitch Fix to match outfits to customers, understand what products have more affinity in which locations, and to replenish and reorder the inventory as to where it should be going.
However, according to June, it is important for the data science team to remember that there is always some aspect of fashion that cannot be predicted. Hence, it is crucial to have confidence bounds, so that it can influence the performance of the style. It also gives a full picture of how well a style can be predicted.
Watch June Andrew's presentation in the video below to hear the full story behind Stitch Fix's algorithms and the challenges their data science team had to face when building the recommender systems.
Full presentation by June Andrews, Data Science Manager of Style Discovery at Stitch Fix, at DSSVirtual.