Source-linked AI summary
M2m: Imbalanced Classification via Major-to-minor Translation
Jaehyung Kim, Jongheon Jeong, Jinwoo Shin
TL;DR
Class imbalance limits DNN generalization under balanced testing, especially when minority classes contain little information. M2m translates majority samples into synthetic minority samples using another classifier, and improves balanced accuracy across diverse datasets, including extreme imbalance.
Problem
Class-imbalanced datasets make DNNs harder to generalize under balanced testing, while minority-focused re-balancing can overfit when minority information is scarce.
Method
M2m constructs balanced training data by translating majority samples into synthetic minority samples through optimization with an independently trained classifier.
Results
17.1% relative balanced-accuracy improvement over standard training and 9.2% over LDAM occurs on the most severely imbalanced Reuters setting.
Takeaways & Limitations
Majority-sample diversity can help imbalanced training learn more generalizable minority features, including under extreme imbalance.
Takeaways & Limitations
M2m depends on an imperfect auxiliary classifier, so generated samples may retain discriminative features from their original majority class and become unreliable, especially when that class is small.
Abstract
from arXiv · showhide
In most real-world scenarios, labeled training datasets are highly class-imbalanced, where deep neural networks suffer from generalizing to a balanced testing criterion. In this paper, we explore a novel yet simple way to alleviate this issue by augmenting less-frequent classes via translating samples (e.g., images) from more-frequent classes. This simple approach enables a classifier to learn more generalizable features of minority classes, by transferring and leveraging the diversity of the majority information. Our experimental results on a variety of class-imbalanced datasets show that the proposed method improves the generalization on minority classes significantly compared to other existing re-sampling or re-weighting methods. The performance of our method even surpasses those of previous state-of-the-art methods for the imbalanced classification.
1. Introduction
M2m addresses class imbalance by replacing duplicated minority oversamples with synthetic samples translated from majority samples, aiming to exploit richer majority information. Across diverse datasets, it improves balanced test accuracy over prior re-sampling and re-weighting methods, including LDAM.
- Class-imbalanced datasets make standard DNN training harder to generalize, especially when evaluation requires balanced class performance.
- Naive re-balancing can overfit minority classes because few minority samples provide insufficient information.
- SMOTE and related methods augment existing minority samples, but often perform poorly under extreme imbalance with very few minority examples.
- M2m generates minority samples by translating majority samples with an independently trained classifier, rather than augmenting existing minority samples.
- M2m improves sampling quality through an optimization objective, sample rejection criterion, and distribution for selecting majority seeds.
- Across synthetic and real-world imbalanced datasets, M2m significantly improves balanced test accuracy over re-sampling and re-weighting methods and surpasses LDAM.For the most severe Reuters imbalance, balanced accuracy improves relatively by 17.1% over standard training and 9.2% over LDAM.
2. M2m: Major-to-minor translation
M2m addresses class imbalance by translating majority samples into synthetic minority samples using a separate classifier, then incorporating them into balanced training. The method combines translation, rejection, and seed-sampling choices to extend minority decision boundaries while leveraging majority diversity.
- Overview: M2m constructs a virtually balanced dataset by adding synthetic minority samples translated from relatively majority samples.The generated samples are added to the training data on the fly rather than produced by augmenting existing minority samples.
- Overview: A pre-trained classifier g guides translation by maximizing the target minority confidence while penalizing confidence in the original majority class.The resulting sample is labeled as the target minority class and used to train the target classifier f.
- Overview: The translation objective seeks a synthetic minority x∗ from a majority seed x0 while preserving majority-class performance through the regularization term λ · f_k0(x).This is intended to teach f minority features captured by g and extend the target minority decision boundary.
- Underlying intuition: Although neural-network translation often produces samples close to the majority seed, the authors report that M2m still improves minority-class generalization.The paper relates these near-seed results to adversarial examples and hypothesizes that majority diversity and minority information captured by g explain the effect.
- Detailed components: M2m rejects generated samples probabilistically, with rejection increasing when the source and target class counts are closer or the source classifier is less reliable.The parameter β controls the modeled reliability of g, and the criterion is motivated by the effective-number-of-samples heuristic.
- Detailed components: Seed-class sampling chooses source classes using an acceptance-probability distribution designed to favor reliable generations while retaining class diversity.After selecting a source class, M2m samples a seed uniformly from that class and generates the required number of samples for each minority class.
3. Experiments
The experiments evaluate M2m across synthetic and naturally imbalanced datasets, showing consistent gains in balanced accuracy and analyzing the method through ablations and feature visualizations.
- Main results: M2m consistently improves balanced accuracy over re-sampling and re-weighting baselines and surpasses LDAM+DRW on the tested long-tailed CIFAR datasets.For CIFAR-LT-10 with N1/NK = 100, M2m reaches 78.3±0.16 bACC versus 77.1±0.49 for LDAM+DRW.
- Main results: M2m performs best among the compared baselines on four naturally imbalanced datasets and is particularly effective on the extremely imbalanced Reuters dataset.The authors describe this as evidence of wider applicability beyond image classification.
- Ablation study: Increasing the diversity of majority-class seed pools progressively improves minority-class accuracy, supporting the role of majority-sample diversity in reducing over-fitting.This trend is reported in the seed-sample ablation.
- Ablation study: Using clean seeds instead of adversarially perturbed translated samples substantially reduces balanced accuracy, indicating that the perturbations are crucial to M2m under the tested setup.The comparison is between original M2m and the M2m-Clean ablation.
4. Conclusion
The paper concludes that Major-to-minor Translation (M2m) uses majority-sample diversity to improve imbalanced learning and suggests broader translation methods as future work.
- M2m is a new over-sampling method for imbalanced classification.
- Majority-sample diversity can help class-imbalanced training, even with translation using a pre-trained classifier.
- The paper suggests exploring more powerful translation methods, such as CycleGAN, in future research.
- The findings suggest adversarial perturbations could be useful features for imbalanced learning when minority classes suffer from insufficient data.
Supplementary Material
The supplementary material is identified only by the paper title in the supplied passage.
- The paper is titled “M2m: Imbalanced Classification via Major-to-minor Translation.”
A. Details on the datasets
The experiments evaluate M2m across image and text datasets, using varied class counts, input representations, model architectures, and dataset splits.
- CIFAR-LT-10/100 contain 60,000 RGB 32 × 32 images across 10 and 100 classes, with 50,000 training and 10,000 testing images.
- CelebA-5 converts CelebA into five-way hair-color classification using blonde, black, bald, brown, and gray labels.
- SUN397 provides scene categorization across 397 classes, using center patches resized to 32×32 and fixed validation and test samples per class.
- Twitter evaluates 23-class part-of-speech tagging with 14,614 training samples and 50-dimensional token embeddings.
- Reuters evaluates text categorization on 36 classes with 6,436 training samples represented as 1000-dimensional bag-of-words vectors.
B. More results from ablation study
The ablation study examines classifier choice, multiple generation classifiers, rejection criteria, and thresholding, with tables reporting repeated-trial performance comparisons.
- M2m uses an independently trained classifier g to generate synthetic minority samples rather than using the training classifier f.
- Table 5 compares ablation performance using means and standard deviations across three random trials, while Table 6 reports ImageNet-LT performance similarly.
- M2m-Ensemble tests whether using two generation classifiers improves generation quality over a single classifier.
- M2m-No-Reject performs significantly worse than M2m when all generated samples are used, supporting the rejection policy.
- Removing the γ threshold causes performance degradation, indicating that the generation classifier’s confidence affects the training classifier.
C. Results on ImageNet-LT
The method is evaluated on ImageNet-LT, a synthetically imbalanced ImageNet subset with 1,000 categories and a Pareto-distributed class imbalance. The evaluation uses standard image augmentations and 128×128 inputs, with the full class distribution shown in Figure 7.
- ImageNet-LT contains 115,846 training samples across 1,000 categories with a Pareto imbalance of α = 6.
- Class sizes range from 1,280 images in the maximal class to 5 images in the minimal class.
- Evaluation preprocessing applies randomly resized cropping, horizontal flipping, and resizing all images to 128×128.
- Figure 7 presents the detailed class distribution of ImageNet-LT.