TopoGR: Preserving Semantic Structure in Generative Recommendation
The newly introduced framework, TopoGR, tackles the structural discrepancy between tokenization and generation in generative recommendation based on semantic IDs. Current approaches regard semantic IDs as separate discrete symbols, failing to consider the topology of the semantic ID space that has been learned. This oversight reduces item relatedness to mere semantic ID overlap, complicating the identification of semantically similar items lacking overlapping IDs. TopoGR employs Bit-decomposable Semantic ID (Binary SID) to maintain the underlying structure. Each Binary SID is developed in a bit-decomposable manner, allowing for deterministic decomposition. This research is documented in a paper available on arXiv under ID 2607.25216.
Key facts
- TopoGR is a topology-preserving generative recommendation framework.
- It uses Bit-decomposable Semantic ID (Binary SID).
- Existing methods overlook the topology of semantic ID space.
- Structural mismatch between tokenization and generation is identified.
- Item relatedness is reduced to exact semantic ID overlap in current methods.
- Binary SID is learned in a bit-decomposable form.
- The paper is available on arXiv with ID 2607.25216.
- TopoGR aims to identify semantically similar items without overlapping IDs.
Entities
Institutions
- arXiv