Is Machine Learning Revolutionizing the Fashion Industry?
What could smart apps do for fashion customers? Can brands and retailers tap into the innovation potential offered by machine learning (ML) and artificial intelligence (AI)? And how?
By Eric Bowman, VP Engineering, Zalando
Algorithms are already being developed today that predict what customers will want to wear before they know it themselves. These algorithms enable accurate predictions of style and taste, combining consumer preferences with current trends captured and processed by powerful ML-based software. The creative minds of the fashion industries will work closely with these intelligent, data-based systems in the future.
The perfect artificial intelligence
ML and AI are not truly helpful to consumers until they know the individual user and their fashion tastes. What inspires a customer's style? What media influence him? Who are his fashion role models? In addition, the software needs to understand the evolution of individual style. What fashion style did the customer like in the past? How has his fashion taste evolved?
Machines learn from data. Data that companies like Zalando and Google (thanks to technologies like TensorFlow) are available and conceal insights into the basic characteristics of all fashion facets. They now enable machines, for example, to recognize and categorize almost any fashion item. To recognize a particular blouse, for example, computers are trained to identify garments that look similar based on thousands or even millions of images. This knowledge is then linked to create complex item representations that reveal symmetries, structures and trends from all areas of the fashion industry.
Matching fashion recommendations are very personal. To truly customize them, a machine must not only understand the fashion world, but also the person. Machines can suggest popular clothing items with relative ease, but that alone is not enough to make a customer happy. A good stylist doesn't just consider the latest fashion trends. Understanding about the particular person, their motivation, interests, dreams and fears is essential for the perfect recommendation.
Machines are already gaining surprising insights from people's surfing and shopping behavior. The next step in personalization, however, still requires the involvement of a human being - machines are not quite that far along today. Even the data and computers of the entire world are not yet able to really know and understand which outfit makes the customer's heart beat faster.
The human factor remains important
How can the stylist logic be automated? The customer's personal preferences are continuously upgraded, taking into account taste, physique, desired style and experience with other consumers. Good stylists bring expertise and people skills to the table. They build a relationship with their customers and help them find the fashion they like - both within and outside the comfort zone of their own or usual style. Brands also play an important role in this, as the designer's talent and personality are expressed in their creations, which thereby appeal to the viewer on an emotional level. The brand's reputation, on which the customer relationship is built, is based on this effect.
The special skills of stylists and designers make them an important part of the fashion world. The human factor is still essential. Not without reason are curated shopping services like Zalon increasingly accepted.
So far, software can't automate it. However, it can identify a customer's style likes and dislikes, sift through data on the latest trends, and find items the customer might like. Stylists can both guide and benefit from machine learning through the use of computers. Software helps stylists, and stylists help optimize the software. Companies like Edited already provide databases for this type of collaboration.
Access to personal fashion vocabulary
Machines learn from many different data points to make the most personalized recommendations possible. Previous purchases and returns from customers are useful, but can also be misleading. When did customers buy for themselves, when was it a gift? Did they return items because they didn't like them or because they didn't fit? In both cases, there are clues the software can use to gain more insight. For example, if the customer is asked if they would like items gift wrapped, that's a great service for them and at the same time helps the software identify the intent behind the purchase. When customers return an item, they should be prompted to provide reasons, so the software can learn more about the customer's tastes or preferred fit.
Customers' surfing behavior also offers insights. Putting all these data points together makes it clearer what is liked and what is not. A kind of fashion vocabulary is created that grows organically as trends change. The software can learn this vocabulary, but it can also help customers with their personal fashion vocabulary. After all, consumers often find it difficult to describe the garment they want. Software can leverage knowledge of customer preferences and behavior so that even quick, ambiguous search queries lead to good results. The software learns not only a general vocabulary for fashion, but a vocabulary set that fits the customer in each case.
Precisely because it's difficult to use the right fashion vocabulary, customers often find other ways to express themselves: they share content, download it, or click "like. So your fashion vocabulary includes more than words: it integrates images, styles, and behaviors.
We want to design software that allows machines to learn from data and from humans to create highly sophisticated applications with a human component. Our team is aware that the human factor is essential to make the connection between fashion and people.
We are still a long way from true artificial intelligence, but by leveraging software along with the creativity and depth of human experience, we can democratize the fashion experience in a whole new way.
Eric Bowman is VP of Technology at Zalando and was a driving force behind the adoption of REST, an API-based approach to software development, and autonomous teams. He has been programming since 1981 and has experienced numerous systems in that time - including Sims 1.0, the first global 3G phone network, TomTom applications, and Gilt Groupe's end-to-end architecture. He is passionate about Scala and continues to contribute code to open source projects.
