Etsy: Building An Algorithm With An Eye For Fashion

Etsy is an online marketplace that connects sellers of handmade and vintage products with buyers from around the world. Etsy was founded in 2005 in a Brooklyn apartment by Rob Kalin, Chris Maguire and Haim Schoppik. Within just two years, Etsy grew to 450,000 registered users and generated annual sales of $26 million. As of 2020, Etsy has 39.4 million active buyers worldwide and 4.3 million sellers. Etsy uses big data to help customers find products easier and to learn about their customer’s shopping habits across the world. 

The Problem With Being One-Of-A-Kind

75% of its 60 million items offered at Etsy are handmade and are therefore one of a kind. Because these items are not mass-produced, they are not easily categorized. A quick search on Etsy for “straw bags” would return 180,000 results showing straw bags of different sizes, styles, materials and prices, everything from monogrammed straw bag to french market straw bags to shoulder messenger straw bags. The options are limitless. 

Applying standard search technology to Etsy’s rich products yields generic results. The standard search algorithm is not built to account for numerous nuances for a single type of item. To make things worse, given the uniqueness of each item on Etsy, it is impossible to divine from search terms and product information listings the exact reasons a product spoke to its customers. 

Pathways to a Just Digital Future

Watch this tech inequality series featuring scholars, practitioners, & activists

Using Big Data To Learn Style 

Etsy turned to big data to solve its search issue. It sought to improve its search experience by matching shoppers to 1) the product they were looking for and 2) the product that best matches their aesthetic preferences. The logic was simple enough – if shoppers love what they saw, then it would increase the chances of purchase, which would subsequently increase buyers and sellers alike as a result of network effects. However, it’s extremely hard to get matching right in a marketplace that caters for shoppers who are looking for unique / bespoke / exclusive finds. More critically, this meant that Etsy would need to teach its algorithm how to understand style – the aesthetics of how a shopper expresses himself / herself. It’s a tall order for an algorithm as it involves understanding different layers of self-identity.

Etsy started using data to train its machine-learning model to effectively identify item styles on site using both textual and visual cues. Fortunately enough for Etsy, it didn’t have to start from scratch. Etsy is well-known in the retail world for its wide data adoption. In fact, 80% of their staff use data on a weekly basis to drive decision making. Staffers poured through data to understand what were people making, what were people buying and what was being sold on their site. They finally identified 42 different styles of goods that had momentum on Etsy. Some examples of these style descriptions include  “boho”, “celestial”, “romantic”, “nautical” etc. Etsy’s data science team then proceeded to collect one month’s worth of search queries, all three million data points, and fed it into their machine learning model. 

In a typical big-time retailer, most AI challenges can be easily solved by an off-the-shelf ML solution. Etsy’s problem stuck out from the rest. Etsy has a large number of small-time sellers who aren’t experts at conveying accurate product description in text. Using ML to scan text alone produced mediocre results that weren’t enough to build an outstanding shopper experience. Etsy found a solution by combining text analysis and image recognition. By combining through millions of images, Etsy’s algorithm is able to conclude that a dress with pineapple print is “tropical” regardless of whether its product description uses this term. This presents a huge opportunity for Etsy to direct shoppers to a wider range of products that the algorithm identifies as “tropical”. 

In order to validate their algorithm, Etsy tested it on even more data. They found that stuff they identified as “tropical” peaked in the summer months and those labelled as “romantic” increased during Valentine’s. This was reassuring for the data science team. 

Etsy’s data science team is still working on perfecting this algorithm. The success of this algorithm would give Etsy an edge over its competitors. The U.S creative products industry alone is valued at $43 billion. E-commerce players such as Amazon are also capitalizing on this growing demand of artisan goods by launching “Amazon Handmade”. However, in the world of analytics, sometimes big data isn’t big enough to meet the demands of machine learning. With 60 million different kinds of unique products, ML would require more than just billions of data points to effectively learn a style.

[1] CNBC Television Mad Money, “Focus on Craftmanship,” [https://www.youtube.com/watch?time_continue=50&v=H-FC2XNqNY4], accessed November 11, 2018.

[2] New York Times, “Inside the Revolution at Etsy,” [https://www.nytimes.com/2017/11/25/business/etsy-josh-silverman.html], accessed November 12, 2018.

[3] Harvard Business Review, “How Predictive AI Will Change Shopping,” [https://hbr.org/2016/11/how-predictive-ai-will-change-shopping], accessed November 11, 2018.

[4] Tech Crunch, “Etsy Buys Blackbird AI to Bring More Machine Learning Into its Search Platform,” [https://techcrunch.com/2016/09/19/etsy-buys-blackbird-ai-to-bring-more-machine-learning-into-its-search-platform/], accessed November 10, 2018.

[5] Forbes, “No More of the Same,”

[https://www.forbes.com/sites/andriacheng/2018/08/08/your-rebellion-against-mass-produced-goods-promises-growth-for-etsy-amazon-handmade/#2377f86327e2], access November 11, 2018.

[6] Venture Beat, “Etsy Opens Machine Learning Center in Toronto,” [https://venturebeat.com/2018/05/17/etsy-opens-machine-learning-center-in-toronto/], accessed November 10, 2018.

[7] New York Times, “Inside the Revolution at Etsy,” [https://www.nytimes.com/2017/11/25/business/etsy-josh-silverman.html], accessed November 12, 2018.