Training with Missed Targets in Generative Recommendation: Separating Supervision from Probability Competition

2026-10-07

AI Summary

This study addresses the probability competition issue in generative recommendation, where appending training objectives to mitigate missed detections degrades ranking performance. We propose a framework that decouples supervision signals from probability competition by constructing a matching loss function to isolate the influence of training objectives on inference candidates, and designing a group-normalized intermediate loss to eliminate competition between inference-only appended objectives and inference candidates. Experiments based on the OneRec model using Amazon datasets demonstrate that our approach improves FT-NDCG by 7.8%–22.2%. These results validate the necessity of evaluating completion strategies according to generator behavior and establish a new paradigm for training optimization in generative recommendation systems.

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.

Problem

• Generative Recommendation
• Missed Targets
• Probability Competition
• Training-Inference Mismatch
• Reranking

Innovation

• Generative Recommendation
• Missed Targets
• Probability Competition
• Matched Losses
• Candidate Completion