Deep Multi-task Learning and Real-time Personalization for Closeup Recommendations

Haomiao Li, Travis Ebesu, Fan Jiang, Jay Adams, Olafur Gudmundsson, Yan Sun, Huizhong Duan13 Jun 2023

Problem

Pinterest's Closeup surface (Related Pins) shows a feed of Pins under any opened Pin and produces the largest share of impressions among its recommendation surfaces. Users respond in several ways (repin, closeup, click, long click), and a shared-bottom multi-task network lets these tasks harm each other. Two retrains with the same configuration can also give very different predictions, which reorders the feed for no reason. Blend weights that combine the predictions had to be re-tuned online after every shift in score distribution, with a long feedback loop.

Example from the article: on Closeup, repin is the most important task, and blend weights must trade off organic engagement against shopping conversion.

Solution in brief

One deep network predicts four user actions through a Multi-gate Mixture of Experts (MMoE). Score regularization toward the production model keeps retrains stable, and transformer-encoded recent actions add real-time context. A second model then learns personalized blend weights from logged randomized weights and a reward, replacing hand-tuned ones.

Questions this case answers
  • How to train one ranking model on several correlated feedback signals without the conflicting tasks hurting each other
  • How to keep predictions of a periodically retrained ranker consistent with the model already in production
  • How to learn per-request weights for combining several predicted outcomes when the hand-tuned weights go stale after every model change
  • How to rank a feed when users can respond in several ways (repin, click, long click) and the tasks conflict?

Tools and methods

  • Multi-task prediction — repin, closeup, click, long click
  • MMoE — experts and per-task gates instead of a shared bottom
  • Score regularization — L2 pull toward the frozen production model
  • Transformer encoder — last 100 actions and 100 engaged Pins
  • Learned Utility — per-request blend weights learned from randomized logs
Continue reading — it's free

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