SCAD: Super-Class-Aware Debiasing for Long-Tailed Semi-Supervised Learning

Abstract

In long-tailed semi-supervised learning (LTSSL), pseudo-labeling often creates a vicious cycle of bias amplification. Recent methods attempt to mitigate this issue via logit adjustment (LA). However, LA-based debiasing remains inherently hierarchy-agnostic and fails to account for semantic relationships between classes. We reveal a critical yet overlooked problem of \textit{intra-super-class imbalance}, where semantically similar classes within a super-class are both highly confusable and locally imbalanced. This combination reinforces early mistakes, causing minority-class representations to be suppressed by their majority neighbors. To break this cycle, we propose Super-Class-Aware Debiasing (SCAD), a framework that performs dynamic, super-class-aware logit adjustment. SCAD leverages latent semantic structure to concentrate its corrective power on the most confusable groups, thereby resolving local imbalances. Extensive experiments demonstrate that SCAD achieves state-of-the-art performance.