0x00 한

Research / Daily reports

  • Sources read: 1
  • Canon files updated: 1
  • Radar entries: 1
In one linePinterest turned on multi-epoch training for its home feed ranking model by freezing the embedding table in the first epoch, and reported a 1.09% gain in online saves.

1Training the home feed ranker for multiple epochs without overfitting — Pinterest

Source Safe and Efficient Multi-Epoch Embedding Training — Pinterest Engineering, 2026-11-06, blog post.

Background Recommendation ranking models are usually trained for 1 epoch, a single pass over the training data. A huge embedding table that holds a vector per item or user ID overfits sharply when it sees the same data twice (multi-epoch overfitting). To use more data, the training window has to be extended, which mixes in older data. This post narrows the cause of this overfitting in the Pinterest home feed ranking model to the embedding table, and reports the results of applying multi-epoch training in production by delaying when the table is trained.

What we learnedA. Freezing the embedding table in the first epoch turns multi-epoch training into a gain

  • When training of the home feed ranking model went from 1 epoch to 2 epochs, the offline save metric dropped by nearly 6% verified — Pinterest blog post.
    • This model had been fine-tuned for 1 epoch on a 3-week window, starting from a foundation model backbone pretrained on data from 5 surfaces.
  • Freezing parts of the model one at a time showed that the embedding table was the main cause of the drop verified — same post.
    • Freezing the table throughout and training for 3 epochs gave +1.25% against the 1-epoch model.
  • The adopted method is a delayed update: freeze the table in the first epoch, and train it together with the rest of the model from the next epoch on verified — same post.
    • On a 10-day window, freezing it through the last epoch gave +0.49%, while training it jointly only in the first epoch and freezing it in the rest gave −0.80%.
    • On a 20-day window, the same approach gave +1.16%, and the gain saturated around 3 epochs.
  • Pinterest turned on multi-epoch training in the production home feed ranking model and got +1.09% online saves verified — same post.

"The modeling and system insights allowed us to enable multi-epoch training in production, leading to +1.09% online saves." — Safe and Efficient Multi-Epoch Embedding Training (blog), 2026-11-06

What we learnedOther changes

  • It writes that in Pinterest's own ads-side research, the ID embedding table did not show this multi-epoch drop verified — same post.
  • On training time, distributed embedding lookup took more than 15% of total training time, so Pinterest added a semi-synchronous pipeline that prefetches the embeddings of the next batch verified — same post.

Canon update Added a Pinterest entry to ## New this month in radar.md.

Why it matters This case separates, in production, the 1-epoch practice of recommendation rankers as being caused by the embedding table. There is no evidence that the tracked services (X, Meta, Netflix, Reddit) use the same training schedule, so it stays on the radar.

Radar

What to check next