AsymRec: New framework beats generative recommenders by 15.8%
Researchers crack the dual bottleneck in generative recommendation with asymmetric encoding.
A new paper from Bin Huang, Xin Wang, and collaborators (including teams from Junwei Pan, Yifeng Zhou, and Wenwu Zhu) tackles two key bottlenecks in generative recommendation (GenRec). Traditional GenRec models represent items as discrete Semantic IDs used symmetrically as both inputs and prediction targets. This creates an "input bottleneck" where lossy quantization loses fine-grained semantics and popularity bias skews representations toward frequent items, and an "output bottleneck" where imprecise discrete targets limit supervision quality.
Their solution, AsymRec, breaks the symmetry by using an asymmetric continuous-discrete framework. Multi-expert Semantic Projection (MSP) maps continuous item embeddings into the Transformer's hidden space via specialized projections, preserving semantic richness and improving generalization for infrequent items. Multi-faceted Hierarchical Quantization (MHQ) builds structured discrete targets through multi-view, multi-level quantization with semantic regularization, avoiding dimensional collapse while keeping fine-grained distinctions. Extensive experiments show AsymRec consistently outperforms state-of-the-art generative recommenders by an average of 15.8%. Code release is planned.
- AsymRec decouples input and output representations to solve the dual information bottleneck in generative recommendation.
- Multi-expert Semantic Projection (MSP) preserves fine-grained semantics and improves recommendation for long-tail items.
- Multi-faceted Hierarchical Quantization (MHQ) creates high-capacity discrete targets, preventing dimensional collapse and boosting accuracy by 15.8% over SOTA.
Why It Matters
AsymRec promises more accurate, fair recommendations that handle rare items better, improving user experience in large-scale systems.