Safe and Efficient Multi-Epoch Embedding Training

This post details Pinterest's successful strategy for enabling multi-epoch fine-tuning of large ID embedding tables in ranking models, overcoming the 'one-epoch phenomenon' that previously led to sharp quality degradation and GPU waste. The core innovation is a decoupled optimization approach: freezing the embedding table for an initial training period (e.g., one epoch) and then allowing it to adapt in later epochs. This strategy was validated through experiments isolating gradient paths and by testing different update timings and data window sizes. The findings demonstrate that the timing of embedding adaptation is crucial, with delayed adaptation yielding significant performance improvements (+1.25% gain over one-epoch model in one experiment) and enabling components like VLMs to receive more optimization time without compromising overall model quality.