Adaptive Resampling-based Training (ART)

• Adaptive Resampling-based Training (ART) is a dynamic technique that adjusts training distributions based on class-level macro F1 scores.
• It blends class frequency with performance-driven weights to shift focus toward underperforming classes during iterative training.
• Empirical results demonstrate that ART consistently outperforms static resampling methods across diverse datasets while being computationally efficient and robust to hyperparameter variations.

Adaptive Resampling-based Training (ART) is a dynamic resampling method for imbalanced classification in which the training-data distribution is periodically updated according to the model’s class-wise performance, rather than being fixed before training begins. In the formulation introduced in "ART: Adaptive Resampling-based Training for Imbalanced Classification" (Basandrai et al., 31 Aug 2025), ART uses class-wise macro F1 scores, computed at fixed intervals, to determine the degree of resampling to be performed. The method is explicitly positioned against traditional resampling procedures that use fixed distributions, such as undersampling the majority or oversampling the minority, and against instance-level difficulty modeling, which is described as noisy and outlier-sensitive. Its defining idea is class-level adaptation: the model incrementally shifts its attention toward underperforming classes in a way that better aligns with the optimization objective (Basandrai et al., 31 Aug 2025).

1. Problem setting and conceptual definition

ART addresses the standard imbalanced-classification setting in which minority classes are underrepresented relative to majority classes, producing systematic bias toward the latter under conventional empirical-risk minimization. The method is motivated by a specific limitation of traditional imbalance remedies: most resampling strategies statically alter the class distribution before or at the start of training and do not respond to changes in class-wise learning difficulty as optimization proceeds (Basandrai et al., 31 Aug 2025).

Within this framing, ART is neither a conventional oversampling method nor a conventional undersampling method. It does not rely on a fixed minority-to-majority ratio, and it does not assume that class frequency alone is an adequate proxy for class difficulty. Instead, it treats per-class macro F1 as an online signal of which classes the model is currently underperforming on, and uses that signal to update the sampling distribution at user-defined intervals. Unlike instance-level hardness schemes such as curriculum learning or online hard example mining, the ART paper emphasizes that class-level adaptation is more stable because it avoids dependence on individual outliers and noisy per-sample difficulty estimates (Basandrai et al., 31 Aug 2025).

A common misconception is that ART is simply another version of SMOTE or random oversampling. The paper’s description does not support that interpretation. ART does not add synthetic points; rather, it emphasizes real samples from classes whose current F1 scores indicate weaker performance. In that sense, ART is a training protocol for dynamically steering minibatch composition, not merely a preprocessing transform (Basandrai et al., 31 Aug 2025).

2. Formal mechanism and training pipeline

The ART formulation begins with a labeled training set

$\mathcal{D} = \{(x_n, y_n)\}_{n=1}^{N},$

where $x_n \in \mathcal{X}$, $y_n \in \{0,1,\ldots,K-1\}$, and $K$ is the number of classes. For each class $i$, the empirical prior is defined as

$$\Pi_i = \frac{1}{N} \sum_{n=1}^N \mathbf{1}[y_n = i].$$

The core adaptive quantities are the class-wise macro F1 score $f_i$, the class difficulty score

$s_i = 1 - f_i,$

the normalized difficulty-based sampling weight $w_i$, and the blended sampling distribution $p_i$. The paper specifies

$x_n \in \mathcal{X}$0

where $x_n \in \mathcal{X}$1 is a blending constant and $x_n \in \mathcal{X}$2 is the boosting frequency, measured in epochs (Basandrai et al., 31 Aug 2025).

The training loop is simple and explicitly epoch-based. Initialization sets uniform class weights $x_n \in \mathcal{X}$3, then defines the initial sampling distribution as $x_n \in \mathcal{X}$4. During each epoch, minibatches are drawn according to $x_n \in \mathcal{X}$5, and model parameters are updated by gradient descent on the minibatch loss. Every $x_n \in \mathcal{X}$6 epochs, ART evaluates per-class F1 on a validation set, recomputes $x_n \in \mathcal{X}$7, normalizes the updated $x_n \in \mathcal{X}$8, and refreshes the class-sampling distribution $x_n \in \mathcal{X}$9 (Basandrai et al., 31 Aug 2025).

This mechanism makes ART adaptive in a strictly class-conditional sense. Classes with lower F1 are assigned larger sampling mass during the next training phase; classes whose F1 improves lose relative emphasis. Because the update is periodic rather than continuous, the method is also straightforward to insert into standard supervised training pipelines that already include epoch boundaries and validation passes. The paper presents this as a practical alternative to methods that require redesigning the loss or generating synthetic samples (Basandrai et al., 31 Aug 2025).

3. Relation to static resampling and algorithm-level imbalance methods

The ART paper compares its procedure to several standard families of imbalance handling. Random over-sampling and SMOTE increase the number of minority examples, either by duplication or by synthetic interpolation. Random under-sampling and NearMiss reduce the majority class. Cost-sensitive learning and focal loss modify the training objective rather than the class distribution. OHEM and related instance-level approaches focus on hard examples, but do so at the sample level rather than the class level (Basandrai et al., 31 Aug 2025).

ART’s stated distinction is that it adapts the training distribution according to actual per-class performance, not just prior class counts or fixed costs. This is the central methodological contrast. SMOTE and NearMiss are static transformations; cost-sensitive learning is typically static once the weights are chosen; instance-hardness approaches can be noisy. ART, by contrast, periodically measures performance and then rebalances sampling toward classes with low macro F1. The paper therefore presents ART as a dynamic, class-level, performance-driven resampling strategy (Basandrai et al., 31 Aug 2025).

The same section of the paper also makes an evaluative claim about consistency: unlike existing methods, whose performance varies by task, ART consistently delivers the strongest macro F1. That claim is tied to the reported benchmarks rather than to a general theorem. A cautious interpretation is that the proposed update rule couples the resampling policy directly to the metric of interest, which is macro F1, and therefore better matches the optimization target in imbalanced settings than frequency-based heuristics do. This suggests that ART should be understood less as a one-time rebalance and more as a feedback controller on class exposure during training (Basandrai et al., 31 Aug 2025).

4. Empirical results and statistical evidence

The reported empirical evaluation covers tabular, image, and text tasks, including Pima Indians Diabetes, Yeast, Red Wine Quality, Long-tailed MNIST, and IMDb-Custom. On these datasets, Table 1 reports the following macro-F1 results for ART: $y_n \in \{0,1,\ldots,K-1\}$0 on Pima, $y_n \in \{0,1,\ldots,K-1\}$1 on Yeast, $y_n \in \{0,1,\ldots,K-1\}$2 on Red Wine Quality, $y_n \in \{0,1,\ldots,K-1\}$3 on MNIST-LT, and $y_n \in \{0,1,\ldots,K-1\}$4 on IMDb-Custom (Basandrai et al., 31 Aug 2025).

For tabular datasets, the paper states that ART improves macro F1 by an average of 2.64 percentage points across all tested tabular datasets when compared to training on the original imbalanced data. It further states that these gains are significant under paired t-tests and Wilcoxon tests with $y_n \in \{0,1,\ldots,K-1\}$5, while results on text and image tasks remain favorable. The same experimental section reports that ART consistently outperforms both resampling-based and algorithm-level methods, including SMOTE, NearMiss Undersampling, and Cost-sensitive Learning, on binary as well as multi-class classification tasks with varying degrees of imbalance (Basandrai et al., 31 Aug 2025).

The paper also provides concrete examples of significance tests in the appendix-oriented summary. For Pima, it reports $y_n \in \{0,1,\ldots,K-1\}$6 against the baseline and $y_n \in \{0,1,\ldots,K-1\}$7 against SMOTE. More generally, it states that for Pima, Yeast, and Red Wine Quality, ART often achieves p-values $y_n \in \{0,1,\ldots,K-1\}$8 against traditional approaches. These reported statistics are used to support the claim that the observed gains are not merely seed-specific fluctuations (Basandrai et al., 31 Aug 2025).

The empirical profile presented in the paper is notable because it spans binary and multi-class problems and includes tabular, image, and text modalities. The reported pattern is not that ART dominates by a large margin on every single task, but that it yields the strongest macro F1 consistently across the evaluated settings. That consistency is one of the central claims made for the method (Basandrai et al., 31 Aug 2025).

5. Advantages, robustness claims, and limitations

The paper attributes several practical advantages to ART. First is dynamic adaptation: the method adjusts sampling in response to actual model performance rather than to fixed dataset statistics. Second is class-level focus: it operates at a stable class-wise level and is described as less noisy than instance-level difficulty modeling. Third is hyperparameter robustness: the paper’s ablation summary states that ART is insensitive to its main hyperparameters, $y_n \in \{0,1,\ldots,K-1\}$9 and $K$0. Fourth is broad applicability: the reported results cover binary and multi-class, tabular, image, and text datasets. Fifth is resource efficiency: ART is reported to achieve strong macro F1 even with small model widths (Basandrai et al., 31 Aug 2025).

These advantages are paired with explicit limitations. ART requires an epoch-based training process and periodic validation F1-score computation. The paper notes that this may not apply to non-iterative models such as decision trees and k-NN. It also notes that computing per-class F1 on the validation set every $K$1 epochs incurs computational overhead, which could be nontrivial for very large datasets (Basandrai et al., 31 Aug 2025).

These limitations clarify the intended deployment regime. ART is best matched to iterative gradient-based learners with well-defined epoch boundaries and a validation loop. It is less naturally phrased for one-shot or non-iterative estimators. The method’s simplicity therefore comes with an architectural assumption: that the training system can repeatedly resample minibatches according to an evolving class distribution and can periodically evaluate per-class validation performance (Basandrai et al., 31 Aug 2025).

6. Terminology, scope, and related adaptive-resampling literature

The acronym “ART” is not unique in the arXiv literature. Distinct papers use ART to denote "Adaptive Robust Transfer Learning" (Wang et al., 2023), "Adaptive Regularized Training" (Glandorf et al., 2023), "Adaptive Randomization Test" (Ham et al., 2022), "Art-based Reinforcement Training" for multimodal LLM fine-tuning (Chudoba et al., 10 Jun 2026), and "Adaptive Reparameterized Time" for diffusion sampling (Huang et al., 26 Jan 2026, Huang et al., 2 Jul 2026). In speech enhancement, the closely related idea of target-conditional resampling appears in "NASTAR: Noise Adaptive Speech Enhancement with Target-Conditional Resampling" (Lee et al., 2022). Accordingly, the phrase “Adaptive Resampling-based Training (ART)” refers specifically to the imbalanced-classification method only when cited as (Basandrai et al., 31 Aug 2025).

More broadly, adaptive resampling has emerged as a recurring design pattern across machine learning domains. REDUS, for example, is a resampling technique for centralized and federated IoT networks that prioritizes misclassified samples and excludes redundant data, reducing the number of training samples per epoch and reporting a training time reduction of up to 72.6% with a minimal accuracy loss of only 1.62% on CICIoT2023 (Gad et al., 2 Jul 2025). In a different line of work, adaptive random Fourier features training stabilized by particle-filter-style resampling is proposed for shallow neural networks and image regression, with resampling introduced to stabilize training and reduce sensitivity to parameter choices (Kammonen et al., 2024). This suggests that ART for imbalanced classification belongs to a wider methodological family in which the sampling process itself becomes a trainable or adaptive component of optimization.

Within that broader family, the defining feature of ART (Basandrai et al., 31 Aug 2025) is its decision to adapt at the class level using class-wise macro F1 rather than at the instance level using per-example difficulty. That design choice is the method’s main conceptual contribution and the basis for its reported empirical behavior across imbalanced binary and multi-class classification tasks.