Conditional Memory Enhances Item Representation in Generative Recommendation
A novel approach known as Conditional Memory Enhanced Item Representation (CMEIR) enhances generative recommendation by tackling issues of information loss and ID collisions in semantic identifiers. Generative recommendation (GR) works by autoregressively generating semantic identifiers (SID) to predict target items. Current item-level representations either combine SID-token embeddings into a single vector, leading to increased quantization loss and obscured code relationships, or utilize external inputs to enhance semantics while neglecting SID structure. CMEIR incorporates a conditional memory module that captures and retrieves item-specific features according to the SID context, thereby improving representation while maintaining structural integrity. Tests on benchmark datasets indicate that CMEIR surpasses leading GR methods. The research is accessible on arXiv.
Key facts
- arXiv:2605.11447
- Generative recommendation (GR) predicts target items by autoregressively generating semantic identifiers (SID).
- Existing item-level representations have two forms: direct merging and external-input-based.
- Direct merging amplifies information loss from quantization and ID collision.
- External-input-based methods cannot reliably preserve SID-structured evidence.
- CMEIR uses a conditional memory module to enhance item representation.
- The method outperforms state-of-the-art GR methods on benchmarks.
- The paper is a cross submission on arXiv.
Entities
Institutions
- arXiv