FSGR framework cuts token frequency bias, boosting recommendation fairness by 20%

Semantic ID (SID)-based generative recommendation has become a powerful approach for turning user-item interactions into token sequences that large language models can Process. However, a new paper from Nankai University reveals a critical flaw: token frequency bias. During SID construction, imbalanced semantic codebooks over-represent popular items, and during training, maximum likelihood estimation amplifies this popularity bias. The result is that high-frequency SID tokens are systematically over-predicted, while low-frequency tokens—often representing niche or emerging items—are barely surfaced. This leads to unfair exposure across item categories and hurts content diversity.

To combat this, the FSGR framework introduces a two-pronged solution. In the SID construction stage, it applies optimal transport (OT)-based assignment optimization and a dual-criteria re-anchor mechanism to create a more balanced semantic representation space. In the recommendation training stage, FSGR adopts a two-stage strategy with hierarchical frequency calibration, which fine-tunes fairness at each layer of the model. The researchers tested FSGR on three public datasets using three backbone models. Results show an average Gini fairness improvement of over 20% while maintaining competitive recommendation accuracy, proving that fairness doesn't have to come at the cost of performance.