How Pinterest Leverages Realtime User Actions in Recommendation to Boost Homefeed Engagement Volume

Xue Xia, Neng Gu, Dhruvil Deven Badani, Andrew Zhai4 Nov 2022

Problem

Pinterest's Homefeed ranker (Pinnability) already used a static, offline-learned user embedding (PinnerFormer) that captures long-term interest. It reacts slowly to what the user is doing right now, which hurts most for new, casual and resurrected users with little history. The team wanted more repins (saves, the main engagement signal) and fewer hides, while keeping the Pinnability architecture and adding only the encoder part of a sequence model.

Illustrative example: a user repins several pins on one topic and expects related pins higher in the same session, while the daily embedding still shows the old mix.

Solution in brief

Feed the ranker the user's latest 100 actions (pin embedding, action type, timestamp) as a sequence. A transformer encoder processes the sequence together with the candidate pin (early fusion). Retrain twice a week to stop the gain from decaying, and move serving to GPU to keep latency flat.

Questions this case answers
  • How to add short-term behavioral intent to a ranking model that already uses long-term user embeddings
  • How to keep the gain of a time-sensitive model from decaying while data and behavior keep shifting
  • How to serve a much heavier ranking model under the same latency and cost budget
  • How to make a feed ranking model react to what a user did in the last minutes, not only to long-term interests?

Tools and methods

  • Realtime action sequence — latest 100 engaged pins with action type and timestamp, built by a streaming pipeline
  • Transformer encoder with early fusion — attends over the sequence with the candidate pin attached to every position
  • Random time window mask — hides actions from the last day during training to avoid a diversity drop
  • Deep & Cross Network (DCN v2) — explicit feature crossing over the full feature set (v1.1)
  • PinnerFormer — existing long-term user embedding, kept alongside the sequence
Continue reading — it's free

Google or a one-time email code. No password.