When Julie Bornstein was vice president of e-commerce at Nordstrom in the early 2000s, someone would manually categorise every item sold online with descriptors: “wrap”, “dress”, “wear to work”. Categorising items this way was slow, costly and imprecise. Fashion is subjective, and product descriptions are open to interpretation.
In the years since, taxonomy — the science of naming, describing and classifying items — has been transformed by data science. Artificial intelligence can help retailers assign more comprehensive descriptions to products, resulting in more lucrative product recommendations. It also significantly increases the speed, scale and complexity of items retailers can onboard. But despite the eventual benefit of automation, establishing these systems requires significant investment from retailers.
Companies have popped up or expanded their reach to usher in the technology. New e-commerce platform Psykhe makes recommendations based on personality traits by identifying both the user and the products; its models can assign products a “personality profile”, informed by traits such as openness or neuroticism, in addition to traditional details, without human input. Resale platform Rebag has developed a universal taxonomy for designer handbags to better appraise products. And Facebook recently unveiled GrokNet, a tool that automatically identifies and describes items in pictures to help people sell items on its marketplace.
How it works
When Bornstein was hiring for her new company, personalised shopping app The Yes, she hired machine learning engineers from Google, Amazon and Facebook in addition to two fashion taxonomists and industry veterans including former magazine editor Taylor Tomasi Hill (as creative and fashion director) and Google’s former head of luxury fashion Lisa Green (as SVP of brand partnerships).
Multiple people at The Yes are involved in developing and maintaining its taxonomy, which aims to include every element of a garment, including physical attributes such as neckline, body style, colour, pattern, cut, product dimensions, material and details such as occasion, season, price and brand. First, the fashion experts provide multiple examples of nuanced attributes such as “French girl style” or “date night”, in addition to details on construction and fit, such as “twill” or “billowy”.
The Yes prioritises and ranks items from its assortment of 60,000 pieces for each user. Each garment has more than 300 attributes; while some overlap, others are different.
© The Yes, artwork by Vogue Business
Data scientists then train the algorithms to identify these characteristics in new products, using a combination of natural language processing to “read” product descriptions (supplied by the brands) and computer vision to “see” product images. The Yes automatically applies attributes to each item, up to more than 300 over time. The Yes co-founder and CTO Amit Aggarwal refers to this as “mapping the DNA of a product”.
If the system flags a product with a low level of confidence, meaning that it is unsure if the attributes it attached are a perfect match, the taxonomists can supplement the machine. If a new type of product or attribute hits the app, such as “Zoom top”, they can add additional attributes. “Taxonomy is always growing. It’s a living, breathing set of information,” Bornstein says. “Even if we’re not seeing something, the model is going to pick up a new set of features that we’re starting to see in repetition.”
Why machines matter
Farfetch’s product catalogue has more than 3,400 brands, and its customer base is global, meaning an intelligent taxonomy is important for creating “credibility at scale”, says Farfetch head of inspiration Natalie Varma, who is responsible for ensuring Farfetch’s tech provides a great user experience. “In fashion, there are so many ways of describing the same thing, which is quite nuanced and is quite internal to the industry, but can be a problem for our customers,” she says. For example, what Rick Owens calls a “duvet coat” might be a “down jacket” at Balenciaga and a “puffer coat” for Prada.
Customers might use many different words to describe the same piece, so comprehensive, specific attributes help recommendation engines make associations between related garments. In 2019, Farfetch categorised approximately 370,000 SKUs.
© Farfetch, artwork by Vogue Business
Garments that arrive at Farfetch are described, photographed and “editorialised” manually; the product’s taxonomy is then enriched automatically using a fashion knowledge graph created by Farfetch data scientists working with its fashion experts. It stores thousands of descriptive fashion terms that have relationships associated with them, which helps in the product recommendation process. Varma says that while there are still some traditional fashion taxonomists on the team, that function is increasingly merging with the duties of data scientists.
In addition to improving fashion search results, AI-enhanced taxonomy can better identify products on a spectrum, rather than in binary terms, Aggarwal says. For example, a human might say an item either is or isn’t a “cold shoulder” top, while a machine can identify it as 80 per cent “cold shoulder”, due to smaller sleeve openings. Machines are also better able to make associations that humans can’t, which offers an alternative to the common e-commerce process of collaborative filtering — an enhanced approach to saying, “We saw you liked this, you might also like that,” he says.
Tech-informed taxonomy can also help in resale. Unlike sneakers, which include standardised style codes, luxury brands do not use standardised or public taxonomy. This makes online comparison-shopping more challenging, says Rebag founder and CEO Charles Gorra. Its appraisal software uses AI to assign handbags a universal “Clair Code” that identifies designer, model, style and size for a database that has 10,000 items. For example, the Clair Code for a Louis Vuitton Palm Springs Backpack PM is HB.LV.PASP.MNCA.PM, which stands for handbag, Louis Vuitton, Palm Springs, monogram canvas and petit modele.
Making long-term progress
In the immediate term, sophisticated taxonomies mean consumers will be better able to find what they want online, with less time and effort. But it will take considerable investment to get there, particularly for retailers that are less advanced in areas like data science and AI. At Facebook, a dedicated AI research team is working to make any image on the platform shoppable, similar to Pinterest, which would take this technology outside of traditional e-commerce product pages. Facebook research scientist manager Tamara Berg says this is a major challenge, but doing so could ultimately help businesses and shoppers.
Psykhe recommends items based on personality characteristics. Similar to The Yes, it invites customers to confirm or remove recommended items.
© Psykhe, artwork by Vogue Business
Bornstein anticipates that after trying to better understand products, retailers will invest in more ways for consumers to provide feedback on the algorithm. However, she acknowledges that many companies might not be equipped to build proprietary systems and that the software-as-a-service platforms that she evaluated early on were not able to perform the functions that The Yes required. She and Aggarwal also say that integrating the 150 brand catalogues, including a drop-ship model, was a “much more massive undertaking” than they originally thought.
Going forward, tech companies specialising in taxonomy hope to expand use-cases. Neuropsychologist and Psykhe founder Anabel Maldonado wants to let people shop according to a Big 5 personality test score, which has a proven relationship with style preferences. Because there are more than 3,000 possible test results, this was a problem that only AI could solve, Maldonado says. Already, retailers to sign on include Moda Operandi, 24Sevres and 11 Honoré, with backing by luxury fashion investor Carmen Busquets.
Both Maldonado and Bornstein think fashion is relatively behind compared to the music and film industries. “Platforms have said, ‘We’ve hired tons of data scientists, we’re sitting on a goldmine of data.’ But you’re thinking, ‘Where is the personalisation?’” Maldonado says, highlighting the need for scientists and industry experts to collaborate. “You need to be asking the right questions.”
To receive the Vogue Business Technology Edit, sign up here.
Comments, questions or feedback? Email us at firstname.lastname@example.org.