Sampling Control for Imbalanced Calibration in Semi-Supervised Learning

TL;DR

This work addresses CISSL under unlabeled-data distribution mismatch by introducing SC-SSL, which decouples sampling control during training from post-hoc logit calibration at inference. A shared-backbone framework employs an expansive classifier and an optimization-bias vector $\mathbf{b}_{opt}$ to expand minority-class decision boundaries and directly adjust logits, grounded by expansion and separation theoretical assumptions. Empirically, SC-SSL achieves state-of-the-art results on CIFAR10-LT, CIFAR100-LT, STL10-LT, and ImageNet-127 across diverse unlabeled distributions, with strong evidence from t-SNE and confusion-matrix analyses showing improved tail-class discrimination. The approach provides practical benefits for real-world CISSL tasks where unlabeled data distributions are unknown or nonstationary, by robustly improving both feature learning and logit calibration via principled sampling control.

Abstract

Class imbalance remains a critical challenge in semi-supervised learning (SSL), especially when distributional mismatches between labeled and unlabeled data lead to biased classification. Although existing methods address this issue by adjusting logits based on the estimated class distribution of unlabeled data, they often handle model imbalance in a coarse-grained manner, conflating data imbalance with bias arising from varying class-specific learning difficulties. To address this issue, we propose a unified framework, SC-SSL, which suppresses model bias through decoupled sampling control. During training, we identify the key variables for sampling control under ideal conditions. By introducing a classifier with explicit expansion capability and adaptively adjusting sampling probabilities across different data distributions, SC-SSL mitigates feature-level imbalance for minority classes. In the inference phase, we further analyze the weight imbalance of the linear classifier and apply post-hoc sampling control with an optimization bias vector to directly calibrate the logits. Extensive experiments across various benchmark datasets and distribution settings validate the consistency and state-of-the-art performance of SC-SSL.

Figures (11)

• Figure 1: We select one image from a head class and one from a non-head class, and visualize their features using Grad-CAM Selvaraju2016GradCAMVE. The top row shows the results from training with uniform sampling, while the bottom row shows the results from SC-SSL. It can be observed that the attention regions for the non-head class are significantly improved, capturing more relevant features. To enhance visual clarity, we applied upsampling.
• Figure 2: Illustration of the proposed framework. Details of the three key factors $\gamma_u$, $\Delta p$, and $\rho$ can be found in \ref{['eq:prob']}. The utilization of these factors can be found in the Method section.
• Figure 3: We choose the distribution setting of Consist to illustrate the pattern of the linear layer's bias term. Under random sampling, since the data follows a long-tailed distribution, the bias term tends to favor head classes. In contrast, under SC-SSL's controlled expansive sampling, the pseudo-label sampling probability is deliberately skewed toward minority classes, resulting in the opposite trend. Therefore, under balanced sampling, the bias term reflects the effect of optimization imbalance.
• Figure : Uniform
• Figure : Uniform

Theorems & Definitions (3)

• Theorem 3.1
• Theorem
• proof