FSGR: Mitigating Token Frequency Bias for Fair SID-Based Generative Recommendation

Semantic ID (SID)-based generative recommendation represents items as sequences of discrete semantic tokens, offering strong generalization and efficient retrieval. However, existing methods suffer from a previously overlooked phenomenon termed Token Frequency Bias, where high-frequency SID tokens are systematically over-predicted while low-frequency tokens are under-predicted. This bias stems from imbalanced semantic codebooks during SID construction, training-data popularity bias, and the maximum likelihood estimation objective. Although various debiasing techniques exist for natural language generation, applying them directly to SID-based recommendation yields suboptimal results because SID tokens exhibit a hierarchical semantic structure that requires layer-specific granularity calibration.

To address token frequency bias, the authors propose FSGR (Fairness optimization framework for SID-based Generative Recommendation), which operates across both SID construction and recommendation training. During SID construction, FSGR employs Balanced Semantic Quantization (BSQ) using Optimal Transport-based Assignment Optimization (OTA) and a Dual-Criteria Re-anchor (DCR) mechanism to encourage uniform codebook utilization and repair geometric voids in the latent space. During recommendation training, FSGR decouples optimization into a two-stage strategy: standard cross-entropy pre-training establishes semantic alignment, and subsequent Hierarchical Frequency Calibration (HFC) applies progressive debiasing across layers according to their semantic granularity.

Evaluated across three public Amazon datasets ("Luxury Beauty", "Industrial and Scientific", and "Software") using three backbone models, FSGR successfully mitigates token frequency bias. Experimental results demonstrate that FSGR delivers an average Gini fairness improvement of over 20% while maintaining competitive recommendation accuracy, proving that joint optimization of semantic quantization and hierarchical frequency calibration effectively improves item-side fairness in generative recommendation.