TRACER reassigns tokens to erase concepts in recommendation systems

Generative recommendation models predict next items by autoregressively generating semantic ID (SID) sequences from user history, making them structurally akin to large language models (LLMs). As privacy regulations tighten, these systems need to unlearn sensitive concepts (e.g., violent content, illegal items) from their training data. However, existing LLM unlearning techniques are ineffective here because SIDs are abstract and often shared by both items to be forgotten and those to be retained, causing severe conflicts that degrade recommendation accuracy.

TRACER (Token ReAssignment for Concept ERasure) introduces an end-to-end framework that reassigns tokens linked to forget concepts to alternative tokens that facilitate forgetting while minimizing side effects on retained items. It employs a coherence regularizer to preserve semantic consistency among retained items after reassignment. Evaluated on real-world recommendation datasets, TRACER effectively removes target concepts while maintaining substantially higher recommendation utility compared to existing unlearning baselines. This work advances practical privacy protection in next-gen recommender systems.