Etsy Details Multi-Task Recommendation Ranker
Etsy's engineering team published a breakdown of its new multi-task canonical ranker for recommendations. The system unifies various recommendation tasks, such as the home feed and search, into a single model. This approach aims to reduce technical debt and create a more consistent user experience across different product surfaces.
- The ranker's architecture is built on a Multi-gate Mixture-of-Experts (MMoE) framework, which replaced older, single-purpose gradient-boosted decision tree models. This structure uses multiple "expert" sub-networks and a gating mechanism that learns how to combine the experts' outputs for different tasks, allowing the model to simultaneously optimize for multiple objectives like predicting favorites and purchases. - To handle various recommendation surfaces (or "modules") within the single model, engineers introduced a `module_name` feature that is fed into each layer of the neural network. This allows the gating network to learn how to weight and combine the expert models differently depending on whether the user is on the homepage, an item page, or another surface, a key step in preventing negative transfer between unrelated tasks. - The initial A/B tests of the canonical ranker showed a 12.5% improvement in "favorite" NDCG (a measure of ranking quality) and also resulted in significant lifts in purchase rates and other engagement metrics. This success demonstrated the model's ability to generalize even to modules whose data was not used during its training. - This unified model is a core part of Etsy's MLOps strategy to reduce technical debt; maintaining hundreds of individual rankers was becoming prohibitively expensive and slowed down the iteration and deployment of new features. The company's ML platform is built on Google Cloud Platform, using tools like Vertex AI and Kubernetes to manage the training, deployment, and real-time inference for these large-scale models. - Etsy's broader AI strategy now includes using generative AI to enhance the structured data for its massive inventory. By using large language models to analyze user activity, the company creates detailed buyer profiles that capture nuanced aesthetic preferences, which can then be fed back into the recommendation and search systems. - The investment in machine learning is a key driver of the company's business strategy, with a stated goal of improving product discovery to grow Gross Merchandise Sales (GMS). In the fourth quarter of 2025, Etsy reported consolidated GMS of $3.6 billion and a record quarterly revenue of $882 million, with executives highlighting that over 75% of push notifications and emails are now personalized. - This multi-task learning approach is similar to architectures at other major tech companies. Netflix, for instance, has moved towards hierarchical multi-task learning to predict user intent as a primary task to inform the secondary task of what content to recommend next. - The engineering team actively monitors for and mitigates challenges common to complex ML deployments. To manage the high-volume, "bursty" nature of search and recommendation requests, they developed an internal tool called Caliper for automated load testing and performance profiling early in the development cycle to address latency issues.
