ComeIR Framework Fixes Identity-Structure Conflicts in Generative Recommendation
New memory-enhanced approach preserves token-level granularity during item generation
Generative recommendation (GR) predicts items by autoregressively generating semantic identifiers (SIDs). Existing methods merge SID-token embeddings into compact vectors or enrich representations with external inputs, but they suffer from two critical issues: the identity-structure preservation conflict (losing SID code relations) and input-output granularity mismatch (token-level generation vs. compressed inputs). This undermines accuracy and reliability.
To address this, the authors introduce ComeIR, a Conditional Memory enhanced Item Representation framework. It reconstructs SID-token embeddings into item-aware inputs and restores token granularity during decoding. Key innovations include MM-guided token scoring (adaptive estimation of each code's contribution), dual-level Engram memory (capturing intra-item composition and inter-item transitions), and a memory-restoring prediction head. Experiments demonstrate ComeIR's effectiveness and flexibility, with scalable gains from enlarged memory.
- Identifies two core problems: identity-structure preservation conflict and input-output granularity mismatch in generative recommendation
- Combines MM-guided token scoring, dual-level Engram memory, and a memory-restoring prediction head for accurate item representation
- Experiments show consistent performance improvements and scalability when increasing conditional memory capacity
Why It Matters
Boosts recommendation accuracy by preserving item identity structure, enabling more reliable and scalable generative systems.