[Paper Review] M2m: Imbalanced Classification via Major-to-minor Translation

Jaehyung Kim, Jongheon Jeong+1 more|arXiv (Cornell University)|Apr 1, 2020

Imbalanced Data Classification Techniques43 references18 citations

TL;DR

This paper proposes M2m, a novel data augmentation method for imbalanced classification that generates synthetic minority-class samples by translating majority-class samples using a pre-trained classifier. By leveraging diverse majority features through adversarial-like perturbations, M2m significantly improves minority class generalization, outperforming state-of-the-art methods like LDAM by up to 17.1% in balanced accuracy on extreme long-tailed datasets such as Reuters.

ABSTRACT

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.

Motivation & Objective

• To address the challenge of poor generalization in deep neural networks trained on highly imbalanced datasets, especially for minority classes.
• To overcome the limitations of existing re-sampling and re-weighting methods that fail under extreme class imbalance due to insufficient minority data.
• To explore whether majority-class diversity can be leveraged to generate more generalizable features for minority classes.
• To develop a simple yet effective over-sampling strategy that avoids overfitting by not relying on existing minority samples for augmentation.

Proposed method

• M2m generates synthetic minority samples by optimizing a majority sample through adversarial perturbations to make it classified as a target minority class by a pre-trained classifier.
• It introduces a novel optimization objective that ensures the translated sample is close to the original while being correctly classified as the minority class, preserving the majority class's performance.
• A sample rejection criterion is designed to favor translation from more frequent majority classes, improving the quality and diversity of generated samples.
• An optimal sampling distribution is proposed for selecting majority seeds to translate, based on class frequency and rejection criteria.
• The method uses a two-stage training setup: first pre-train a classifier on the imbalanced dataset, then use it to guide the translation of majority samples into synthetic minority samples.
• The synthetic samples are used to over-sample the minority class during training, improving generalization without overfitting.

Figure 1: An overview of the proposed method, called Major-to-minor Translation (M2m). M2m is based on the over-sampling method, and attempts to replace the over-sampled (duplicated) minority samples with synthetic ones translated from other majority samples. The more details are presented in Sectio

Experimental results

Research questions

• RQ1 Can translating majority-class samples into minority-class representations improve generalization on minority classes in imbalanced learning?
• RQ2 Does leveraging the diversity of majority samples through translation lead to better feature learning than traditional minority over-sampling methods?
• RQ3 How does the use of adversarial-like perturbations in sample translation affect model generalization and robustness?
• RQ4 Can a method that avoids using existing minority samples still achieve state-of-the-art performance on extreme long-tailed datasets?
• RQ5 Is the performance gain from M2m due to the adversarial nature of the generated samples, or is it due to the diversity of the majority features?

Key findings

• M2m significantly improves balanced accuracy on all tested datasets, including CIFAR-LT-10, CIFAR-LT-100, ImageNet, CelebA, SUN397, Twitter, and Reuters, compared to standard ERM and re-sampling/re-weighting baselines.
• On the Reuters dataset with extreme class imbalance, M2m improves balanced accuracy by 17.1% over standard training and 9.2% over LDAM, a state-of-the-art margin-based method.
• The ablation study shows that using clean majority samples (M2m-Clean) instead of adversarially perturbed ones leads to a significant drop in performance, confirming the importance of the perturbations.
• t-SNE visualization confirms that M2m produces more separable and well-distributed feature representations for minority classes compared to other methods.
• The cumulative false positive curve for M2m is consistently below other methods and closer to linear, indicating more uniform and lower error distribution across all classes.
• The method outperforms SMOTE and other SMOTE variants in low-data regimes, demonstrating its effectiveness when minority classes have very few samples.

Figure 2: An illustration of M2m generation. A majority seed $x_{0}$ is translated to a synthetic minority $x^{*}$ based on the decision boundary of $g$ . By incorporating $x^{*}$ , $f$ learns an extended decision boundary of the target minority class.