• Etsy serves 100M+ listings, using recommendation modules to guide buyers at every shopping stage. • Each module follows a two‑phase pipeline: fast candidate selection then ML‑based ranking. • Traditional approach trained a separate ranker per module on its own implicit feedback. • The new multi‑task canonical ranker shares a single model across all modules, leveraging pooled data. • Features include user context (recent purchases, clicked categories) and item attributes (title, taxonomy). • Optimized for click‑through and conversion rates, the model improves relevance and scalability across Etsy.

Article Summaries:

  • Etsy, which hosts over 100 million listings, relies on recommendation modules to surface relevant items at different stages of a shopper’s journey. Historically each module had its own ranker, trained on that module’s implicit feedback. As the number of modules grew into the hundreds, maintaining separate models became costly and slowed innovation. To address this, Etsy introduced “canonical rankers” - shared models optimized for a single engagement metric (e.g., click‑through or conversion) but designed to power multiple modules. The first canonical ranker targets visit frequency, aiming to surface items that encourage future returns while matching or exceeding the performance of the former one‑to‑one rankers.
  • Etsy, which hosts over 100 million listings, relies on hundreds of recommendation modules (“modules”) that surface items to users at different shopping stages. Traditionally each module had its own ranker-an ML model that scores candidate items for relevance-trained only on data from that module. As the number of modules grew, maintaining separate rankers became costly and slowed innovation. To address this, Etsy introduced “canonical rankers”: shared models optimized for a single engagement metric (e.g., click‑through or conversion) but designed to serve multiple modules. The first canonical ranker targets visit frequency, aiming to surface items that encourage users to return. This approach promises comparable performance to module‑specific rankers while reducing engineering overhead and enabling faster feature rollout.
  • Etsy, which hosts over 100 million listings, relies on recommendation modules to surface relevant items at every stage of a shopper’s journey. Historically each module had its own ranker, trained on that module’s implicit feedback, which became costly to maintain as the number of modules grew. To reduce engineering overhead and keep recommendation quality high, Etsy introduced “canonical rankers” - models optimized for a single engagement metric but designed to serve multiple modules. The first canonical ranker targets visit frequency, aiming to surface items that encourage users to return. By sharing a single model across modules, Etsy expects comparable performance with lower training and maintenance costs.
  • Etsy, which hosts over 100 million listings, relies on recommendation modules to surface relevant items across its web and mobile platforms. Historically each module used a dedicated ranker trained on its own implicit‑feedback data, a strategy that scaled poorly as the number of modules grew into the hundreds. Maintaining separate models became costly and slowed experimentation. To address this, Etsy introduced “canonical rankers” - multi‑task models optimized for a single engagement metric (e.g., click‑through or conversion) but designed to serve multiple recommendation modules. The first canonical ranker targets visit frequency, aiming to surface items that encourage repeat visits while matching or exceeding the performance of module‑specific rankers, thereby reducing engineering overhead and improving scalability.
  • Etsy has moved from a one‑to‑one recommendation system-where each of its hundreds of recommendation modules used a dedicated ranker-to a unified “canonical ranker” approach. The new rankers are trained on data from multiple modules and optimized for the same engagement metrics (e.g., click‑through or conversion rates), enabling them to serve several recommendation tasks simultaneously. This shift reduces engineering overhead, lowers training and maintenance costs, and allows rapid incorporation of new features. Etsy’s first canonical ranker targets visit frequency, aiming to surface items that encourage users to return and explore more of the platform.

Sources: