Generative recommenders improve ranking by separating training supervision
Training with Missed Targets in Generative Recommendation: Separating Supervision from Probability Competition
Information RetrievalMachine Learning
Summary
Generative recommendation systems try to suggest items to users but sometimes miss items they should have recommended. The authors study how adding these missed items during training affects the final ranked suggestions. They find that making missed items compete with original candidates for ranking can actually harm performance. By separating supervision on missed items during training from the ranking competition during inference, recommendation quality improves. This means recommendation systems should carefully choose how they train on missed items rather than always using the same approach.
What this means in practice
- •For e-commerce recommendation teams: Improve product ranking by training recommendation models without forcing missed targets to compete with original candidates during inference.
- •For online video platform engineers: Enhance video recommendation lists by selectively applying appended-target training based on performance gains per generator model.
Authors
Xuesi Wang, Yangbin Shi, Xiaolin Zheng
Abstract
Generative recommenders return a limited candidate set and may omit observed targets before reranking. A training strategy appends these missed targets to reranker training lists, although inference still ranks only original candidates. This operation simultaneously changes retrieved-target weight, adds supervision over appended targets, and makes the two groups compete for probability. An append/no-append comparison therefore cannot explain changes in returned-item rankings. We construct three matched losses that hold retrieved-target weight fixed while introducing appended-target supervision and group competition separately. The intermediate loss trains within both groups but normalizes them separately, preventing training-only targets from competing with inference candidates. Experiments with a released OneRec model and locally trained Amazon generators show that this competition can harm returned-item ranking. In four prespecified Amazon Video Games comparisons, removing it improved full-target normalized discounted cumulative gain (FT-NDCG) by 7.8--22.2\%; 95\% intervals over users and three of four intervals over training runs excluded zero. A conservative development-set rule selected appended-target training for two of three generators in one held-out category and rejected it for all three in another, avoiding a 1.7\% loss. Candidate completion should therefore be evaluated for each generator rather than applied automatically.