While human inspiration will always be the primary source of new fashions, Sylvia Chang discovers how machines are learning to spot current trends and predict new ones.

When Thailand's King Bhumibol Adulyadej died in October 2016, a fashion brand in Hong Kong targeting mid- to low-price range customers sold out their black and white colored shirts - the colors of respect worn during the official mourning period for the late king - in almost one night. Designers and merchandisers couldn't help but wonder: what if we can be informed of big events beforehand and predict which colors will stir up the next shopping spree accordingly?

Traditionally, industry experts predict fashion trends based on their own experiences and sense of fashion. These days, however, social media comes into play with its impact on consumers' behavior. And with the help of artificial intelligence technologies, a revolution in the fashion industry is underway.

"We intend to provide a real-time estimation of fashion color trends by big data from social media," said Gloria Yao, director of project development at the Hong Kong Research Institute of Textiles and Apparel. Yao and her team are working on a model of color prediction that uses AI technologies.

Yao said sales prediction is very important to a fashion business as it facilitates effective inventory management, reduces markdowns due to overstocking and thereby increases profits.

Social media impact

According to a 2016 report by US-based advertising agency PMX, social media posts drive 6.3 percent of website traffic to luxury brands. Among all social media platforms, Facebook drives the most sales in fashion.

Under this phenomenon, Yao and her team collected hundreds of influential Facebook accounts in the fashion industry. These include over 500 fashion brands, 100 fashion magazines, 150 designers and 500 social media celebrities. To be qualified as influential, Yao said, a brand should have at least 100,000 followers, a magazine with 10,000 followers and a designer or a celebrity with 1,000 followers. In addition, all these accounts should have frequent updates and be able to raise conservations in their communities.

"The aim is to capture the ongoing social and cultural events from social media posts and to validate their impact on consumers' preference on color," Yao said.

A technology in AI, Natural Language Processing, is applied to recognize the texts and images of these Facebook posts. It helps to identify "authentic fashion posts" that relate to color. By this, Yao means posts that really talk about color in fashion.

She explained, "If a celebrity says (on a post), 'I'm walking on a red carpet in a film festival,' the 'red' here has nothing to do with fashion. But if she says, 'I like my red-colored dress today,' then this 'red' is the target of our data collection."

HKRITA works with a well-known local fashion brand which provides its data on sales, inventories, prices, shop locations and marketing information.

By comparing the data of Facebook posts and of the brand, the team found a match: the color trend started on Facebook runs ahead of the real trend of selling.

"If we can further find out the time gap between the two trends and their relations, we will have more accurate color predictions," Yao said. Historical data reveals an accuracy of about 70 percent. This may change as data are updating on a rolling basis.

The whole project will finish in April 2019 when a license of production will be introduced into the industry and commercialized for application.

Until then, the model of color prediction will be mature to fit different brands, Yao said, "but with some adjustments." The reason is that different brands may target different consumers and therefore, show different color preferences.

Applying AI to makeup

More and more fashion brands have applied AI technologies. Cosmetics giant L'Oreal allows customers to "try on" lipsticks, eye shadow and eyelash styles in a real setting of their smart phones. Luxury brands like Burberry give shoppers intelligent advice on which fashion items are most suited for them.

"AI enhancements will go beyond the traditional areas of machine tasks into creative and customer interaction processes, blurring the line between technology and creativity," reads a 2018 report on fashion trends by New York-based consulting firm McKinsey & Company.

Fashion is about design. It's a concept in which physical materials are combined with human sensations through the sense of sight, touch, hearing and smell. All in all, it's about the emotions and feelings of a customer. How can this sentimental thing be understood by the rational technology of AI? How can fashion be read by a machine?

The first step is to teach a machine to analyze a fashion image. This step is "to make sure the machine understands the fashion world," said Calvin Wong. He's a professor in the Institute of Textiles and Clothing at Hong Kong Polytechnic University. Wong, who is leading a team of a dozen researchers, is cooperating with Alibaba Group to enhance their website search functions on Taobao.

At present, the biggest challenge facing Taobao is the sheer chaos of products that have not been tagged by professions, said Jia Menglei, senior staff engineer at Alibaba Group. Customers complain it's hard to find what they're looking for. With the data set composed of fashion attributes, produced by Wong's team, tagging would become more exact and predictable. "This enables us to rebuild the world map of consumer goods," Jia said.

"There are just tons of items on Taobao. I often have to spend hours on it selecting what I need. Sometimes it felt dizzying," said 26-year-old Li Yuan, an online shopping fantastic.

Understanding a sweater

Apparel comes in innumerable patterns, each with different characteristics and elements that define it as "fashion." The variety of materials, colors, shapes and styles create a world of experience in skirts, trousers, t-shirts, dresses, sweaters and so on, all according to individual taste. A sweater can have various types of collars: Peter Pan collars, puritan collars, shirt collars, rib collars and so on.

To train a machine to understand a sweater, you need to install a data set of all kinds of images of sweaters, tagged with their design attributes. The AI network, or a single machine trained to have predictive ability, would then recognize a specific type of sweater next time it meets the image.

The data set is most important. It serves as a guideline for the machines' learning process and sets the foundation for all follow-up research. "It needs to be professional enough to meet the requirements of machine training," Wong said. By professional, he means the knowledge of fashion including complete and precise attributes of each style that may be applied to the basic images.

"For a sweater with a turtleneck, you need to give it a tag both as a sweater and a turtleneck. If it's identified sometimes as a sweater and sometimes as a turtleneck, the machine will be confused." Wong said at present, the accuracy of searches for clothing images on the internet is only about 50% effective on average. He attributes that to the lack of enough professional data.

The AI system adopted by Alibaba Group has been trained in more than 500,000 outfits by stylists on Taobao. The machine learning has yielded about 30 attributes and over 150 tags, establishing a set of fashion details.

The second level of AI application in fashion is to analyze current styles, evaluate their most popular elements, provide recommendations and even predict fashion trends. As Wong of PolyU said, if a machine has a data set of the latest images of outfits in the fashion world, it could calculate which fashion elements are emerging as popular and may become the fashion trend of the foreseeable future. Then it could offer recommendations for consumers, based on predictive analysis and individual preference.

The team expects that in the near future, AI will be able to collect more data from customers, providing important information about the popularity of certain styles. The data notes the amount of time a customer spends looking at a particular piece of clothing on the smart mirror. It records the amount of time the customer "tries on" the piece, the final decision of purchase and so on.

Alibaba Group will combine data from Taobao with the data collected in offline shops. The company intends to provide an all-round data service for the industry chain, ranging from fashion design and material selection to storage and logistics.

AI appears set to transform the fashion industry. Is there anything a machine is not likely to master in the near future? Yes. In the end, designers are irreplaceable. The core of design, Wong said, is the inspiration, and the machine can only serve as an assistant, at best.

Contact the writer at

sylvia@chinadailyhk.com

(HK Edition 10/05/2018 page4)