AI for search at Etsy

Search Evolution

The evolution of search technology has been marked by significant advancements, particularly with the integration of AI and machine learning. Andrew Stanton explains that early search systems relied on rule-based algorithms, but the introduction of machine learning models like support vector machines in the late 1990s and early 2000s marked a turning point 1. These models improved relevancy by moving from generative to discriminative approaches. Today, search encompasses various areas, including information search, e-commerce search, and question-answering systems, each with unique challenges and applications 2.

We've always tried to use and try to understand relevancy. And so we've had things like TF, IDF and VM 25 which have kind of existed for several decades at this point.

--- Andrew Stanton

The rise of deep neural networks and unsupervised learning has further transformed the search landscape, impacting how search engines operate and improve user experience.

   

Etsy's Strategy

Implementing advanced search solutions in e-commerce requires a strategic approach to optimize relevance and ranking models. Andrew Stanton describes Etsy's unique challenges, such as handling over 60 million results, many of which are handmade or one-of-a-kind items 3. Unlike traditional search engines, Etsy lacks structured data like SKUs, relying on machine learning to parse and rank items effectively. The integration of neuroevolution experiments at Etsy has improved search relevance, but it also highlighted the need for careful metric optimization to avoid unintended outcomes 4.

We put it at the very end, and we call it the kind of the business intelligence layer, and it allows us to kind of incorporate both beliefs or priors about what would be beneficial for the marketplace.

--- Andrew Stanton

Future efforts will focus on refining these strategies and understanding marketplace trade-offs to enhance search performance.

   

Innovative Models

The search industry is witnessing a revolution with the adoption of state-of-the-art models and AI-driven innovations. Andrew Stanton highlights the impact of deep learning and neural IR, which have replaced traditional text-based retrieval systems with massive neural networks 5. These advancements allow for more accurate translation from query to document space, enhancing search precision. The blending of machine learning with distributed systems is becoming more prevalent, as seen in the integration of billion-parameter neural nets in information retrieval stacks by companies like Amazon 6.

The best search systems are really going to incorporate techniques from both worlds into the code that's being built, rather than having them segregated apart.

--- Andrew Stanton

This convergence is driving the development of more sophisticated search solutions that address real-world challenges effectively.