| Abha Bhattarai |
CAN’T decide what to wear?
Uniqlo, the Japanese fast-fashion chain, has a solution: A chatbot that gives clothing recommendations based on human input, as well as your purchasing history and… your horoscope.
The technology, which has been years in the making, is just one example of the extremes that retailers are going to as they try to build computer algorithms that can intuit the intangibles of fashion.
“Instead of making something that’s purely mechanical – you bought this last month, so you might like this – we’re infusing humanity into the process,” said Rei Inamoto, founder of Inamoto & C, the New York-based firm behind Uniqlo’s technology. “When somebody asks, ‘What should I wear?’ they’re looking for a personalised answer.”
As retailers race to offer customisation and convenience, they increasingly turn to stylists and personal shoppers to win over consumers and to help fine-tune algorithms that might give them an edge in the USD3 trillion global fashion industry.
Stitch Fix, the online styling-subscription service, has assembled about 3,700 remote stylists – from stay-at-home moms to full-time lawyers who fancy themselves fashionistas – to select clients’ outfits based on a combination of sales data, artificial intelligence and their own taste. (The pay: $15 an hour.)
Meanwhile, tech giant Amazon has hired dozens of fashion designers, photo editors and retail workers in recent years to help shape its proprietary software. The company’s researchers have also developed an algorithm that analyses images of clothing and then designs similar items, according to an MIT Technology Review report. (Jeff Bezos, the founder and chief executive of Amazon, owns The Washington Post.)
“Companies are realising that you can squeeze even more juice out of the orange if you combine data analytics with the human stylists,” said Wendy Liebmann, Chief Executive of consulting firm WSL Strategic Retail. “We all know that artificial intelligence is a valuable tool, but it so often misses the nuances.”
Sarah is in her mid-30s, lives in Massachusetts and works in client services. She’s looking for summer workwear that can transition into the fall.
She likes: polka dots, floral and lace.
She dislikes: stripes, jackets and black.
Rachel Gee, a preschool teacher-turned-Stitch Fix stylist, knew all of this the moment she glanced at Sarah’s profile. The company’s algorithms have distilled Sarah’s body measurements and clothing preferences (which she provided when she signed up for the service), as well as three years’ worth of purchases, into easy-to-read data points. Sarah has long legs and a short torso, and she tends to spend USD50 to USD100 per item.
Gee scrolled through her client’s Pinterest page, which is full of bohemian styles, embroidered details and geometric prints, then checked her Twitter and Instagram accounts for more clues.
“I can see she’s very romantic and edgy, stylewise,” Gee said. “She’s outdoorsy and has a casual vibe. I feel like I know her – like she’s my friend, almost.”
Now, Gee said, she was ready to pick five items to send to her client. The algorithm’s top suggestion was a pair of distressed denim shorts. Not today, Gee said. She sent Sarah a pair of shorts the week before, and, plus, Sarah was looking for officewear.
She found an olive green Calvin Klein dress with a subtle floral print. The computer told her that Sarah loves Calvin Klein and predicted a 51 percent success rate. “We’re more than halfway there,” Gee said. That’s a pretty high probability that she’s going to keep that dress.”
She picked four more items: A navy pencil skirt, a magenta Calvin Klein blouse, an off-white knit blouse and a “fun statement necklace” with stones. Each item cost between USD50 and USD100.
“This whole process is like a partnership between me and the data,” said Gee, 29, who lives in San Francisco and now works full time for Stitch Fix.
Stitch Fix, with its 3,700 stylists, had only 100 five years ago. Executives say they look for workers with a background in fashion, styling, customer service or retail. (They also want someone with “solid writing skills, including grammar and punctuation,” a spokeswoman said, because stylists send personalised notes in each shipment.)
“As we learn more about each client over time, both our algorithms and stylists become more accurate,” said Meredith Dunn, the company’s vice president of styling and client experience. “Our stylists read and digest feedback from clients and our algorithms ingest that data, too.”
The company has also used that data to create 18 private-label brands to fulfil consumer demand for items such as “funky printed” dress shirts and “timeless” first-date wear.
Some in the industry, though, say the model isn’t sustainable. Working with a personal stylist at Bergdorf Goodman or Saks Fifth Avenue is one thing; relying on machine-learning and stylists in far-off cubicles is another, and it seems like a stretch, said Milton Pedraza, chief executive of the Luxury Institute, a market research firm in New York.
“Algorithms and one-size-fits-all stylists keep costs down, but it doesn’t mean that they’re particularly good matchmakers or can understand tastes and lifestyle,” Pedraza said. “Having a stylist is about creating a personal relationship, and that just doesn’t happen if someone is styling you from a computer on the other side of the country.”
Andrea Alder was a year out of fashion school when an Amazon recruiter approached her with a top-secret job offer: To train the company’s machines to become arbiters of style.
For the next year, she spent 40 hours a week, sometimes more, sitting in a cubicle and voting on peoples’ outfits. Her mission was twofold: to provide real-time feedback to consumers who sought Amazon’s advice on what to wear and to teach the company’s algorithms how to assess clothing for fit, trend, silhouette and, eventually, style. (The “outfit compare” feature on Amazon’s app allows Prime members to ask for guidance on which of two outfits looks better. Trying to decide between a blue blouse and a purple one? Amazon promises human judgment within minutes.)
Alder began her shift at 5.30 most mornings, condensing her fashion education and styling experience into split-second decisions. She’d see two photos side by side, and quickly vote on which outfit looked better. There was no time to ask questions (Was this stuffy suit on its way to a corporate job or a wedding?) or to explain her decisions – either to the person she was styling or the algorithm she was training.
“The way the machine learns is through a lot of repetition,” said Alder, 24, who like many of her peers signed a nondisclosure agreement before taking the job. “Everything has to be very simple. It’s like explaining to a little kid why something works or doesn’t.”
Alder found a certain rhythm to her work. Demand for outfit input swelled at certain times, like right before work and in the run-up to New Year’s Eve. She learned to be quick – and deliberate – in her judgments. But she was wary of what her split-second decisions might be teaching Amazon’s machines about fashion and style.
“Machines want black-and-white, and so much of fashion is subtle,” she said. “If you pick a dress over a pair of pants, you don’t want the algorithm to think it’s because dresses always look better than pants. So how do you make it understand that this particular dress fits well and is on trend while the pants are outdated?” – WP-BLOOM