Rebalancing with Calibrated Sub-classes (RCS): A Statistical Fusion-based Framework for Robust Imbalanced Classification across Modalities

TL;DR

This work tackles extreme class imbalance across image, text, and tabular modalities by introducing Rebalancing with Calibrated Sub-classes (RCS), a statistical fusion-based framework that blends majority and intermediate-class statistics to calibrate minority distributions. It combines an autoencoder with supervised, contrastive, and reconstruction objectives to produce a disentangled latent space, and uses Gaussian Mixture Models plus a novel distribution-calibration scheme to generate faithful synthetic minority samples in latent space (then decoded to data space when needed). Key contributions include the three-way class categorization (majority, intermediate, minority), theoretical guarantees for calibrated means and unbiased sampling, and extensive empirical validation showing consistent gains over state-of-the-art oversampling and calibration methods across eight image/text/tabular benchmarks. The results demonstrate robust, scalable improvements in imbalance settings, with practical impact for real-world deployments where minority patterns are subtle and high-dimensional data abounds.

Abstract

Class imbalance, where certain classes have insufficient data, poses a critical challenge for robust classification, often biasing models toward majority classes. Distribution calibration offers a promising avenue to address this by estimating more accurate class distributions. In this work, we propose Rebalancing with Calibrated Sub-classes (RCS) - a novel distribution calibration framework for robust imbalanced classification. RCS aims to fuse statistical information from the majority and intermediate class distributions via a weighted mixture of Gaussian components to estimate minority class parameters more accurately. An encoder-decoder network is trained to preserve structural relationships in imbalanced datasets and prevent feature disentanglement. Post-training, encoder-extracted feature vectors are leveraged to generate synthetic samples guided by the calibrated distributions. This fusion-based calibration effectively mitigates overgeneralization by incorporating neighborhood distribution information rather than relying solely on majority-class statistics. Extensive experiments on diverse image, text, and tabular datasets demonstrate that RCS consistently outperforms several baseline and state-of-the-art methods, highlighting its effectiveness and broad applicability in addressing real-world imbalanced classification challenges.

Figures (9)

• Figure 1: Figures depict the t-SNE plots of feature vectors in the latent space for the MNIST dataset. (a) Shows the balanced feature vectors when the vanilla autoencoder is trained using the balanced data. (b) Displays the balanced feature vectors when our autoencoder is trained using the balanced data. (c) Illustrates the imbalanced feature vectors when trained using the vanilla autoencoder on the imbalanced dataset. (d) Represents the imbalanced feature vectors from our autoencoder network. (e) Visualizes the oversampling from the latent feature vectors obtained from our trained autoencoder network.
• Figure 2: Illustrations of our Data Calibration method, the first figure (from top left) represents the imbalanced feature vectors, the second figure represents the Gaussian Mixture components of the Majority class as well as the Intermediate classes, the third figure represents the oversampling in the latent space, and the visualization of the components participating in estimating the mean and variance corresponding to each point of the minority class from which synthetic feature vectors are sampled from, in the fourth figure corresponding to the estimated mean and variance we sample synthetic feature vectors from this Gaussian distribution and in the fifth figure we display the oversampled feature vectors.
• Figure 3: Illustrations of our method, Stage 1 depicts the autoencoder training, and Stage 2 portrays the data calibration in the latent space, followed by a classification phase where we employ ResNet32 architecture for Classification.
• Figure S1: Figures depicting the performance of our method by varying the hyperparameters. a), e), i) depicting the balance accuracy corresponding to the hyperparameter $\eta,k \text{ and }t$ respectively. b), f), j) depicting the Mathews Correlation Coefficient corresponding to the hyperparameter $\eta,k \text{ and }t$ respectively. c), g), k) depicting the F1-Score corresponding to the hyperparameter $\eta,k \text{ and }t$ respectively. d), h), l) depicting the Gmean corresponding to the hyperparameter $\eta,k \text{ and }t$ respectively
• Figure S2: Illustrations of the majority class samples and the generated samples in the other classes for MNIST, FMNIST, CIFAR10, and SVHN (left to right)

Theorems & Definitions (14)

• Definition 1
• Theorem 3.1
• proof
• Theorem 3.2
• proof
• Remark 2
• Definition 2
• Definition 3
• Theorem 3.3
• proof