Asymmetric GenRec Framework Addresses Dual Bottlenecks
A new asymmetric generative recommendation framework, AsymRec, tackles dual information bottlenecks in existing GenRec models. Traditional GenRec uses discrete Semantic IDs symmetrically as inputs and outputs, causing lossy quantization and popularity bias in input, and imprecise targets in output. AsymRec decouples these representations via Multi-expert Semantic Projection (MSP) for continuous input embeddings and Multi-faceted Hierarchical Quantization (MHQ) for output. MSP preserves semantic richness and improves generalization to infrequent items. The framework is detailed in arXiv:2605.14512.
Key facts
- AsymRec is an asymmetric continuous-discrete framework for generative recommendation.
- It addresses input bottleneck from lossy quantization and popularity bias.
- It addresses output bottleneck from imprecise discrete targets.
- Multi-expert Semantic Projection (MSP) maps continuous embeddings into Transformer hidden space.
- MSP uses expert-specialized projections to preserve semantic richness.
- Multi-faceted Hierarchical Quantization (MHQ) constructs improved discrete targets.
- The paper is available on arXiv with ID 2605.14512.
- The approach decouples input and output representations.
Entities
Institutions
- arXiv