LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing

2025-04-09

AI Summary

In class-imbalanced semi-supervised learning (CISSL), classifier performance on minority classes degrades due to data bias in pseudo-labeling. Method: This paper proposes a bias-mitigation mechanism based on consistency-induced gradient conflict. It introduces (1) a theoretical proof that the black image serves as the optimal baseline for logit subtraction, grounding logit debiasing; (2) a gradient-direction conflict detection module that identifies samples whose pseudo-label gradients oppose those of debiased logits, converting their bias signals into optimization incentives; and (3) a dynamic pseudo-label refinement strategy integrating consistency regularization and baseline calibration for enhanced robustness. Results: Evaluated on multiple CISSL benchmarks, the method significantly improves minority-class accuracy and consistently outperforms existing debiasing approaches.

Technology Category

• Machine Learning: Semi-Supervised Learning
• Computer Vision: Bias, Fairness & Privacy
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Classifiers often learn to be biased corresponding to the class-imbalanced dataset, especially under the semi-supervised learning (SSL) set. While previous work tries to appropriately re-balance the classifiers by subtracting a class-irrelevant image's logit, but lacks a firm theoretical basis. We theoretically analyze why exploiting a baseline image can refine pseudo-labels and prove that the black image is the best choice. We also indicated that as the training process deepens, the pseudo-labels before and after refinement become closer. Based on this observation, we propose a debiasing scheme dubbed LCGC, which Learning from Consistency Gradient Conflicting, by encouraging biased class predictions during training. We intentionally update the pseudo-labels whose gradient conflicts with the debiased logits, representing the optimization direction offered by the over-imbalanced classifier predictions. Then, we debiased the predictions by subtracting the baseline image logits during testing. Extensive experiments demonstrate that LCGC can significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

Problem

• Addressing class-imbalanced bias in semi-supervised learning classifiers
• Theorizing baseline image use for refining pseudo-labels effectively
• Proposing LCGC to debias predictions via gradient conflict learning

Innovation

• Using black image logits for pseudo-label refinement
• Learning from consistency gradient conflicting
• Debiasing by subtracting baseline image logits