Fashion’s one constant — change — forces everyone within the industry to stay ahead of trends in order to make intelligent business decisions. But designers, marketers and buyers may soon be getting help with some of the “grunt work” involved in analyzing data. A study revealed that AI-powered “classifiers” were more reliable at analyzing garments, and could classify footwear styles and footwear subcategories, more accurately and consistently than apparel professionals, according to EDITED.
When compared to humans, the classifiers made, on average:
- 2.5 percentage points fewer errors when identifying garment types;
- 9.3 percentage points fewer errors when determining subcategories; and
- Approximately 6.5 percentage points fewer errors when classifying specific footwear styles.
These results don’t mean, however, that the technology replaces the human touch — it just enhances it. In showing more efficiency at performing redundant tasks such as identifying products and styles, AI-powered solutions could give designers and merchandise buyers more time and energy for developing product-driven strategies and eliminating data-related guesswork.
“If you’re talking about designing a new product, you can filter down the data and see the details of the products,” said Kris Graham, Data Scientist at EDITED in an interview with Retail TouchPoints. “Ultimately, you can see the ‘soft’ colors and shapes that it would be really difficult to narrow down to on your own. If you’re in buying, you can find your assortment for the next year. Even beyond that, it goes down to the supply chains — if you see in the data that there’s a lot of a specific color or a specific type of fabric that’s coming through, then you can supply all those things and you can prepare as well. You can identify things before they become apparent in the market.”
The Classifier Benchmarking test was conducted online. In July 2018, 50 fashion and apparel professionals identified nearly 1,286 products that had been randomly selected from EDITED’s database. Their answers, along with the classifier’s predictions, were compared to each product’s category in order to determine accuracy.
‘Task Fatigue’ Can Plague Humans, But Not AI
Each respondent study identified only 57 products, a sample size designed to prevent participants from thinking that the project was a grueling task. At the same time, it was designed so that the respondents could get tired easily, particularly because they were using a skillset they wouldn’t necessarily dedicate much effort to improving upon.
This “tiredness” factor differentiates machine learning from human analysis — respondents’ accuracy over time dwindled 4%, according to the study. AI doesn’t get tired or experience task fatigue. In fact, it’s designed to improve accuracy as it analyzes more products.
“If you’re going to classify millions of products, it’s going to take you an awfully long time,” said Graham. “You’re going to be tired and your performance may drop over that period. The amount of time it would take you to analyze its product — it’s just not a feasible process to look at every product on the market. That’s really where AI can help retailers to round up all of that raw, messy, noisy information and allow them to make data-based decisions.”
Footwear Categories Tough For Both Humans And Machines
Digging deeper into the study results revealed these findings:
- Classifiers outperform humans on even trivial tasks of identifying garment types, at 97.8% accuracy against 95.4% for humans, and 96.7% accuracy against 87.4% for humans on identifying footwear subcategories;
- Both classifiers and humans struggled to accurately identify footwear styles, at 69% and 63% respectively, suggesting that this categorization is still very complicated within the apparel industry; and
- Classifiers can classify 57 products in a sample dataset in a matter of seconds, while it took, on average, 6.5 minutes for a human respondent to classify them — the equivalent of nearly 2.5 hours for the whole dataset.
The classifier identifies most footwear subcategories well, with the exception of slippers — a small category where only 40% are categorized correctly (the remaining 60% are identified as shoes). Human classifications were scattered across the board: 55% of slippers were labeled as shoes, 21% as sandals, 7% as trainers and 3% as boots.
“The variation in human error was larger, with different types of error,” said Graham. “A human may mistake a trainer for a sandal, and a trainer for a boot, and a trainer for a slipper. The classifier may mistake a slipper for a shoe, but you wouldn’t have five or six types of error.”