synthetic oversampling

To address the weak generalization and severe prediction bias of large language models (LLMs) on few-shot and class-imbalanced text data, this paper proposes an embedding-space synthetic feature augmentation method. Unlike conventional approaches, it operates directly in the language model’s latent embedding space—bypassing raw text generation—and jointly synthesizes minority-class features via embedding interpolation, noise perturbation, and adversarial generation to optimize semantic representation distributions. The method integrates seamlessly into standard fine-tuning pipelines and is compatible with mainstream open-source text classification benchmarks. Experiments across multiple benchmarks demonstrate up to a 12.3% improvement in minority-class F1 score, alongside consistent gains in overall accuracy and robustness. The core innovation lies in migrating synthetic data generation from the input space to the embedding space, enabling efficient, lossless, and fair representation calibration.

This study investigates the mechanism by which synthetic data augmentation improves score-based classification performance—measured by metrics such as AUROC and AUPRC—in class-imbalanced settings. By developing a theoretical framework that disentangles the effects of augmentation on effective class weighting and distributional bias, and integrating tools from statistical learning theory, minimax analysis, and finite-sample error decomposition, the work establishes that under correctly specified models, augmentation solely reduces variance without improving overall performance. However, under model misspecification, it can mitigate ranking errors by correcting class imbalance. The analysis yields novel minimax lower bounds, which are corroborated through simulation experiments.

This work addresses the challenge of label imbalance in multi-label classification, which often biases models toward frequent labels and degrades overall performance. To mitigate this issue, the authors propose a novel oversampling method based on label-specific distances. The approach introduces a label-aware distance metric that dynamically constructs a weighted feature subspace tailored to each label, enabling the selection of label-consistent nearest neighbors for synthetic sample generation. This strategy enhances both the label consistency and boundary representativeness of the synthesized instances. Experimental results across multiple multi-label datasets demonstrate that the proposed method significantly outperforms existing oversampling techniques, effectively alleviating label imbalance and improving classification performance.

In imbalanced classification, scarcity of minority-class samples induces model bias and spurious correlations. Method: This paper proposes a novel synthetic oversampling paradigm leveraging large language models (LLMs), establishing the first theoretical framework for synthetic data in imbalanced learning. It rigorously quantifies performance gains, derives scaling laws linking synthetic sample size to model accuracy, and characterizes the capability boundary of Transformers for generating high-fidelity synthetic samples. Contribution/Results: Theoretically, the method provably enhances classification accuracy, robustness, and generalization. Empirically, LLM-generated samples effectively mitigate class bias and outperform conventional resampling techniques (e.g., SMOTE) across multiple benchmarks. This work delivers an interpretable, scalable, and LLM-driven solution for trustworthy imbalanced learning.

To address sample selection bias arising from the coexistence of class imbalance and label noise in deep active learning, this paper proposes a robust one-dimensional threshold-driven active learning paradigm. The method jointly models class imbalance and label noise—first achieved in active learning—and employs deep feature embedding followed by one-dimensional projection to robustly estimate an inter-class separation threshold. This threshold defines a priority region near the decision boundary where high-uncertainty samples are selected for labeling. The framework is theoretically compatible with batch querying and label-noise tolerance. Evaluated on multiple imbalanced benchmark datasets, it reduces annotation cost by over 60% compared to state-of-the-art active learning methods and improves accuracy by more than 80% relative to random sampling, while significantly enhancing minority-class recognition performance.

This study investigates the efficacy limits and optimal scale of synthetic data augmentation in class-imbalanced learning. By establishing a unified statistical learning framework grounded in balanced population risk analysis, the work reveals that augmentation benefits model performance only under “local asymmetry” conditions, and that the optimal number of synthetic minority samples depends critically on the generator’s accuracy and bias direction—challenging the conventional assumption that perfect class balance is inherently optimal. To address this, the authors propose a validation loss–based strategy for tuning the synthetic sample size (VTSS). Both theoretical analysis and empirical experiments demonstrate that ill-conceived augmentation can degrade performance, whereas VTSS reliably identifies the optimal augmentation scale, with consistent validation on both simulated data and real-world sepsis prediction tasks.

To address model bias arising from severe class imbalance in tabular data—such as in fraud detection and medical diagnosis—where minority-class samples are critically scarce, this paper proposes an AI-driven oversampling method based on synthetic data generation. Unlike conventional oversampling techniques, our approach leverages the open-source MOSTLY AI Synthetic Data SDK to generate high-fidelity, mixed-type synthetic instances that precisely populate sparse regions of the minority class in feature space, thereby enhancing distributional diversity and representativeness. Extensive experiments on multiple real-world imbalanced datasets demonstrate that the proposed method consistently outperforms baseline oversamplers—including SMOTE and ADASYN—across key metrics such as F1-score, AUC, and G-mean. Notably, performance gains are most pronounced under extreme imbalance (minority class < 1%), underscoring the efficacy and generalizability of synthetic-data-driven oversampling for mitigating class bias.

Adaptive label smoothing (ALS) is widely adopted for regularization, yet its efficacy under diverse image augmentations—such as random erasing, noise injection, and TrivialAugment—remains poorly understood, particularly regarding robustness to common corruptions. Method: We extend ALS to operate across multiple heterogeneous augmentation operators and systematically evaluate its regularization behavior and robustness under both single-strong and composite-augmentation regimes. Results: ALS significantly improves generalization under strong single augmentations (e.g., aggressive random erasing), but its benefits vanish—or even become detrimental—under combined augmentations. Moreover, excessive smoothing degrades model robustness to noise, blur, and other common corruptions. This work identifies, for the first time, the failure mechanism of ALS in heterogeneous augmentation settings and establishes a critical balance condition between augmentation strength and diversity for effective ALS deployment. Our findings provide both theoretical insight and practical guidelines for applying ALS in realistic training scenarios.

In class-imbalanced classification, synthesizing minority-class samples often introduces distributional bias, leading to model overfitting and degraded generalization. To address this, we propose a novel framework that estimates and corrects synthesis-induced bias using distributional information from the majority class. Unlike conventional approaches assuming synthesized samples follow the true minority-class distribution, our method leverages structural consistency in majority-class features to construct a provably consistent bias estimator, coupled with dynamic error calibration during training. Theoretically, we derive bounds on the bias estimation error and provide guarantees on improved prediction accuracy. Empirically, extensive experiments on benchmark datasets—including MNIST—demonstrate significant gains in F1-score, AUC, and robustness against label noise. Moreover, the framework naturally extends to multi-task learning and causal inference settings, offering broad applicability without architectural modification.

Medical diagnostic models often suffer from poor generalizability due to class-feature bias and class imbalance—i.e., overreliance on spurious features strongly correlated with only certain classes. To address this, we propose a class-unbiased training framework that jointly tackles both issues: (1) an inter-class inequality loss explicitly enforces balanced feature–class associations by penalizing disparities in feature attribution across classes; and (2) class-weighted distributionally robust optimization (DRO) improves worst-case performance under long-tailed class distributions. Our method is validated on hybrid datasets combining synthetic and real-world multicenter medical data. Experiments across multiple diagnostic tasks demonstrate consistent gains: average accuracy improvements of +3.2–5.8%, and a 37% reduction in performance standard deviation across classes—indicating markedly enhanced cross-class stability. To our knowledge, this is the first work to simultaneously decouple and model class-feature bias and class imbalance within a unified framework.