AI

A Language for Sneakers? GOAT Applies AI in Product Management

Sneakerheads of the world love online reseller GOAT. Sellers submit formatted photos of their sneakers and the company finds a buyer. The GOAT website has some 400,000 listings. Rare models like Air Yeezy Blink, Air Jordan 3 Retro Solefly and Pharrell x Chanel x NMD Human Race Trail can fetch tens of thousands of dollars.

Artificial intelligence is revolutionizing transportation, radiology, even music recommendations. So why not the sneaker reselling market?

Sneakerheads of the world love online reseller GOAT. Sellers submit formatted photos of their sneakers and the company finds a buyer. The GOAT website has some 400,000 listings. Rare models like Air Yeezy Blink, Air Jordan 3 Retro Solefly and Pharrell x Chanel x NMD Human Race Trail can fetch tens of thousands of dollars.

But how to properly classify and cross-reference all these different shoes, or detect counterfeits? To solve these and other problems, GOAT has turned to AI.

Founded in California in 2015, GOAT has 400 employees, eight million users, and a net worth of US$250 million. About 60 of GOAT’s employees are engineers and data scientists, who help the startup leverage AI to tackle various challenges.

GOAT has a large, high-quality sneaker dataset, and the company has applied machine learning in its product catalog management to enhance search and recommendations based on shoe appearance.
goat sneaker.jpg
GOAT senior data scientist Emmanuel Fuentes recently published a post on Medium explaining how the company’s AI systems build their own language for describing the visual characteristics of sneakers.

Fuentes refers to the method as “embeddings”, which essentially involves creating a constrained variational autoencoder (VAE) framework to learn and represent sneakers’ unique styles with diverse latent factors. Such high-dimensional data can be visualized using software such as T-distributed Stochastic Neighbor Embedding (t-SNE), which enables for example adding bulk annotations.

Latent factors in multiple dimensions are relatively independent and these variables enable the study of the relationships between sneakers in the product catalog based on attributes such as sole, upper, silhouette, color, pattern, material, etc. Even two sneakers that look quite different may share style features which the learned embeddings can detect for example to make match recommendations.

anchor.gif
Transitions Between Two Sneakers According to Latent Space

GOAT researchers experimented with arithmetic operations in the latent space. For example, direct addition of two sneakers’ latent variables will produce language for a third sneaker in the existing database that combines visual characteristics of both original shoes.

AI techniques are also used to check whether shoes are fake — which is critical in a premium collectors’ marketplace. Rather than using human experts, GOAT relies on computer vision and machine learning along with instrument and industry knowledge to examine a shoe’s physical attributes and spot counterfeits. Fakes are returned to the sellers, only authentic shoes are sent to buyers.

GOAT is now adopting machine learning across its entire supply chain — including warehouse management and order distribution — in order to improve overall customer experience.

 

Source: BigDataDigest via Synced China


Localization: Tingting Cao | Editor: Michael Sarazen

0 comments on “A Language for Sneakers? GOAT Applies AI in Product Management

Leave a Reply

Your email address will not be published.

%d bloggers like this: