Novel Fuzzy Kernel Learning Framework Improves Classification on Imbalanced Datasets

A newly developed machine learning framework significantly improves classification performance on imbalanced and noise-corrupted datasets by coupling domain-level prior knowledge with intuitionistic fuzzy modeling. Introduced by researchers Brajendra Singh Rajput, Sunita Soni, and Bhagat Singh Raghuwanshi, the method is designated as the Universum Relative Density Intuitionistic Fuzzy Kernelized Weighted Extreme Learning Machine, or URDIFKWELM. Published in the International Journal of Machine Learning and Cybernetics, the study provides a robust mathematical framework designed to overcome the algorithmic biases and decision boundary distortions that frequently degrade automated classification systems in real-world data environments.

Handling imbalanced datasets remains a major practical difficulty across numerous machine learning tasks. As a direct consequence, the resulting decision boundaries become heavily skewed toward the majority distribution, leaving minority class instances with insufficient representation during the training process.

In particular, the universum twin support vector machine (UTSVM) remains susceptible to this fundamental limitation. When trained on disproportionate data distributions, UTSVM frequently produces models dominated by the majority class. Under these skewed conditions, minority samples are often incorrectly treated as noise rather than meaningful pattern signals, thereby sharply reducing predictive accuracy.

To address these dual challenges of class imbalance and data corruption, the authors formulated the URDIFKWELM architecture. The framework draws inspiration from the integration of prior domain knowledge with intuitionistic fuzzy modeling principles, implemented within a kernelized weighted extreme learning machine structure.

A core innovation of the proposed framework is the strategic integration of Universum samples as prior information.

Complementing the Universum data integration, the framework applies a Relative Density Intuitionistic fuzzy membership mechanism. This mechanism is explicitly designed to control the influence of uncertain, noisy, and irregular samples throughout the training routine.

The authors demonstrated the efficacy of URDIFKWELM through extensive experimental evaluations conducted on real-world imbalanced datasets. To thoroughly examine the model’s resilience, the experiments evaluated performance under both clean baseline settings and challenging environments corrupted by Gaussian noise. Across these comprehensive trials, the proposed approach consistently achieved superior performance compared with several state-of-the-art methods. The experimental findings confirm that combining Universum prior data with relative density intuitionistic fuzzy modeling enhances robustness against noise, strengthens minority class modeling, and improves overall generalization performance.

Subject of Research: Classification of imbalanced and noisy datasets using intuitionistic fuzzy extreme learning machines

Article Title: Robust universum relative density intuitionistic fuzzy kernel weighted extreme learning machine for imbalanced data

Article References: Rajput, B. S., Soni, S., & Raghuwanshi, B. S. (2026). Robust universum relative density intuitionistic fuzzy kernel weighted extreme learning machine for imbalanced data. International Journal of Machine Learning and Cybernetics, 17(12), Article 532. https://doi.org/10.1007/s13042-026-03363-3

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03363-3

Keywords: machine learning, imbalanced datasets, extreme learning machine, universum learning, intuitionistic fuzzy modeling, relative density, kernel methods, twin support vector machine, decision boundaries, gaussian noise, noise robustness, minority class modeling, outlier suppression, pattern recognition, fuzzy membership, classification algorithms, generalization performance