class imbalance mitigation

Class imbalance severely biases model training and undermines evaluation validity across domains. Method: This paper systematically reviews 258 authoritative publications (2003–2023) and proposes the first cross-domain, multi-dimensional taxonomy for imbalance learning—unifying sampling-based methods (e.g., SMOTE, ADASYN), cost-sensitive learning, ensemble techniques (e.g., EasyEnsemble, RUSBoost), deep learning adaptations, and evaluation metrics (F1, G-mean, AUC-PR). Contribution/Results: We construct a full-stack knowledge graph spanning preprocessing, modeling, evaluation, and deployment, and introduce the first evaluation selection guideline tailored to large-scale, real-world imbalanced applications. The framework significantly lowers practical adoption barriers in high-skew domains such as financial risk control and medical diagnosis, while identifying emerging research frontiers—including self-supervised and causal learning integration.

Class imbalance severely degrades model discrimination for minority classes, critically hindering deployment in high-stakes domains such as healthcare and finance. This paper systematically surveys over one hundred imbalance mitigation strategies, introducing the first unified taxonomy that integrates generative approaches (e.g., GANs, VAEs) with classical resampling techniques—including SMOTE, neighborhood density estimation, and adaptive threshold-based resampling. We further propose a multidimensional evaluation framework and practical deployment guidelines tailored to real-world constraints. Empirical validation across diverse benchmark tasks demonstrates that the surveyed methods improve minority-class F1-score by 12–35%. Crucially, we identify a novel pathway for jointly optimizing interpretability and generalization—bridging theoretical advances with engineering feasibility. This work provides a comprehensive, actionable foundation for both advancing imbalance learning theory and enabling robust, trustworthy deployment in critical applications.

To address class imbalance in long-tailed multi-class classification—where existing methods (e.g., resampling, cost-sensitive learning, loss modification) lack theoretical grounding and fail to satisfy Bayes consistency—this paper proposes the first learning framework that is both theoretically rigorous and practically effective. Our method introduces: (1) a strongly H-consistent margin-based loss function tailored for class-imbalanced settings; (2) a class-sensitive Rademacher complexity theory, yielding tight generalization error bounds; and (3) the IMMAX algorithm, enabling margin-driven, hypothesis-class-agnostic robust optimization. We provide formal theoretical guarantees establishing strong generalization performance. Empirically, IMMAX achieves statistically significant improvements over state-of-the-art methods across multiple benchmark datasets, validating both its effectiveness and superior generalization capability in long-tailed classification.

Linear classifiers (e.g., SVM) suffer from degraded generalization performance on high-dimensional imbalanced data. Method: We establish a high-dimensional asymptotic theoretical framework and, for the first time, rigorously derive analytical expressions for the generalization error under undersampling and oversampling. Our approach integrates random matrix theory, high-dimensional statistical learning, and unsupervised probabilistic modeling–driven resampling. Contribution/Results: We quantify how resampling efficacy depends on the first- and second-order statistics of the data and the choice of evaluation metric. Crucially, we prove—and empirically verify—that hybrid sampling consistently outperforms either undersampling or oversampling alone. Extensive numerical experiments and evaluations on real-world datasets—including deep neural network features—demonstrate strong agreement between theoretical predictions and empirical results, with substantial improvements in minority-class classification accuracy. This work provides an interpretable, generalizable, and principle-based foundation for data rebalancing in high dimensions.

In noisy label learning, sample selection suffers from dual biases: data bias (imbalanced selection sets) and training bias (error accumulation). To address these issues, this paper proposes ITEM, a noise-tolerant expert model. ITEM is the first to jointly model and mitigate both biases in a unified framework. It introduces a lightweight multi-expert robust network architecture, integrated with a dual-weighted class-discriminative sampler and a hybrid mini-batch training strategy. Additionally, an error-robust optimization mechanism is incorporated to enhance generalization under label noise. Extensive experiments on multiple benchmark datasets with synthetic and real-world label noise demonstrate that ITEM consistently outperforms state-of-the-art methods, achieving average accuracy gains of 3–5% while reducing parameter count by over 20%. The source code is publicly available.

This study addresses the issue of model bias toward majority classes in imbalanced classification by formally framing it as a label shift domain adaptation problem between the source distribution (observed data) and the target distribution (balanced evaluation distribution). The authors introduce the concept of “transfer cost” and provide theoretical analysis showing that SMOTE incurs higher transfer cost than random oversampling methods such as Bootstrap in medium- to high-dimensional spaces. Building on this framework, they integrate minority class distribution estimation into data augmentation and empirically demonstrate that random oversampling generally outperforms SMOTE in such settings. These findings offer both theoretical justification and practical guidance for selecting oversampling strategies in imbalanced classification tasks.

Empirical Risk Minimization (ERM) suffers from degraded generalization under long-tailed class distributions. Method: This paper proposes a data-dependent shrinkage technique and establishes the first fine-grained, class-aware unified generalization upper bound. Unlike conventional coarse-grained analyses relying on global statistics, our bound explicitly quantifies how class-specific terms influence generalization error. Contribution/Results: The bound provides the first systematic theoretical explanation of the intrinsic mechanisms underlying reweighting and logit adjustment—resolving several counterintuitive empirical observations. Leveraging this theory, we design a principled learning algorithm that significantly improves minority-class accuracy on standard long-tailed benchmarks—including CIFAR-10-LT and ImageNet-LT—outperforming state-of-the-art methods.

This work addresses the instability and performance degradation commonly observed during fine-tuning of pre-trained models, which often stems from gradient cancellation leading to optimization collapse. To mitigate this issue, the paper introduces, for the first time in the context of fine-tuning, a dynamic gradient scaling mechanism, proposing the Dynamic Scaled Gradient Descent (DSGD) algorithm. DSGD adaptively attenuates the gradient magnitudes of correctly classified samples, thereby effectively alleviating gradient cancellation. The method substantially enhances fine-tuning stability and robustness, consistently reducing performance variance and achieving higher accuracy than existing approaches across multiple benchmark datasets and large-scale models.

The impact of class imbalance correction on model discriminative performance and probability calibration in clinical risk prediction remains unclear. This study systematically evaluates the effects of SMOTE, random oversampling (ROS), and random undersampling (RUS) across ten real-world clinical datasets using a range of linear and nonlinear models. Comprehensive comparisons are conducted using metrics including ROC-AUC, Brier score, and calibration intercept/slope. Results indicate that none of the three resampling methods significantly improve discrimination, yet all consistently degrade probability calibration—evidenced by increased Brier scores (0.029–0.080) and substantial shifts in calibration parameters—revealing systematic distortion in predicted risk estimates. These findings challenge the conventional use of resampling techniques in clinical prediction modeling.

This work addresses the challenge of accurately predicting deep, rare classes in hierarchical multi-label classification, where such categories suffer from intrinsic low frequency and further diminished prevalence due to hierarchical propagation. To this end, we propose a novel loss function that explicitly focuses on rare nodes—rather than rare samples—by integrating node-level class imbalance weighting with a focal weighting mechanism grounded in ensemble-based uncertainty quantification. This approach dynamically adjusts training emphasis based on model uncertainty and seamlessly integrates into mainstream neural architectures such as CNNs. Extensive experiments demonstrate substantial improvements, with recall gains up to fivefold and significantly higher F₁ scores compared to baseline methods. Notably, the proposed method maintains robust performance even under challenging conditions, including suboptimal encoders or scarce data regimes.

This work addresses class-imbalanced classification in capacity-constrained scenarios—such as rare disease detection and fraud identification—by proposing a novel framework that explicitly embeds user-specified constraints on the proportion of positive predictions directly into the optimization objective. The approach ensures that the total number of selected positive instances never exceeds the prescribed operational capacity while maximizing detection performance. Notably, it achieves precise control over the actual number of positive predictions, overcoming a key limitation of conventional resampling techniques like SMOTE, which cannot regulate the final selection volume. Built upon standard machine learning models, the framework naturally supports both batch and online learning without requiring complex architectural modifications. Experimental results demonstrate that, under strict adherence to capacity constraints, the proposed method significantly outperforms established baselines in classification effectiveness.