Teaching Algorithms to Have Style

As an e-commerce site for artisans, Etsy faces a number of unique data challenges. Not only are a significant percentage handmade and one-at-a-time, the artisans who write product descriptions are regular folks, not merchandising or search engine optimization experts. For example, a quick search on Etsy for “earrings” delivers over 5 million results ranging from ornate $50,000 diamond earrings to cheap stick-on earrings for children to a life-sized bronze statue of a well accessorized horse. 

To help better match shoppers with the right items, Etsy developed a machine learning (ML) algorithm that analyzes a customer’s style based on past purchasing/browsing behavior and, using those insights, ranks search results based on those aesthetic preferences. 

Using a combination of natural language processing and computer vision, they developed a tool that analyzes product pages to classify products by style. For example, the tool might scan the product page of a vintage rattan chair from the 1950’s and categorize it as “mid-century” or “minimalist ” even if the product description does not overtly use those terms. 

However, this was only the beginning. In the year or two since then Etsy has refined their personalization approach considerably. 

Interaction-Based Product Recommendations

There are multiple ways a user can interact with a product listed on Etsy. They can click on it, favorite it, add it to their cart or save it for later, just to name a few. However, some of these interactions are more meaningful than others and not necessarily in the way one expects. 

While traditional personalization methods analyzes what products a customer interacts with, Etsy’s recommendation engine also looks at how the customer interacts with products. As explained in a 2021 publication, Etsy’s new product recommendation model:

• Learns multiple embeddings for each listing to model different user intents, expressed through the different ways in which a user interacts with a listing
• Generalizes upon co-occurrence based models for accurately predicting which listing a user may want to interact with next without falling prey to sparsity
• Incorporates a co-occurrence regularization term that guarantees equal or better performance than baseline co-occurrence models
• Captures co-occurrence patterns in a low-dimensional representation that makes it computationally efficient to compare any two items, allowing for use as a feature or candidate set selection method in production recommender systems
• Provides more interpretable recommendations based on the interaction type of a customer’s previous activity

Another related challenge Etsy faces is that, given its huge array of products, people don’t shop on Etsy for just one product or person. They may buy a wedding present for a friend one day and a halloween costume for themselves the next. Furthermore, while some customers buy products on a regular basis throughout the year, many are either new to the site or only use the site once a year. Ensuring these shoppers, for whom Etsy does not have a wealth of past behavior data to mine, also find what they’re looking for as efficiently as possible is also very important.

One solution to these challenges is a what Esty call an 𝑂nline 𝑃ersonalized 𝐴ttribute-based 𝑅e-ranker (OPAR), a ML algorithm that “learns which attributes the buyer likes and dislikes, forming an interpretable user preference profile and improving re-ranking performance over time, within the same session.” In other words, it delivers product recommendations based on real-time data. 

Personalization Algorithms That Also Boost Revenue

Though Esty’s top priority is to match shoppers with whatever it is they’re looking for, as with any business, they also need to make sure search results are optimized from a revenue perspective as well. However, unlike traditional e-commerce sites, Etsy is a two-sided market place that serves both buyers and sellers. 

As Dr. LiangJie Hong, Etsy’s former Director of Engineering, Data Science and Machine Learning, explained in a 2018 podcast, “Relevance is one way to look at the things, but we want to optimize revenue, which is called gross merchandise value. We’re going to optimize when people search things. It’s not only we want to provide the most revenue result, but also the result that can generate the most revenue. Then we need to model how likely you are going to click on that. And then after you click on that thing, how likely you are going to purchase that thing. And you also need to take the pricing into account. 

Do we recommend things that have a higher conversion rate by the lower price or a low conversion rate but a very high price? You see all these trade-outs and all these compromises that we need to make, such as that we adopt the traditional model to optimization of revenue in the e-commerce area.”

Though creating algorithms that align with conflicting business metrics might seem impossible, Etsy has made significant strides in this endeavor with its Joint Optimization Framework. Though we won’t go into full detail here, the model combines relevance, diversity and market-level metrics to deliver a “method for imposing business needs while eliminating many of the common forms of interventions that lead to sub-par search experiences.”

Become an AI Data Analytics NETWORK member today!

Premium content. Expert insight. Instant access. All for FREE.