FUZZY ENSEMBLE EMBEDDING LEARNING (FEEL): A CLASSIFIER-AGNOSTIC FRAMEWORK FOR UNCERTAINTY-AWARE AND CALIBRATED DECISION-MAKING

Abstract

Uncertainty and vagueness are inherent in real-world data, particularly in domains such as healthcare, finance, and natural language processing. Traditional machine learning models, which rely on crisp features and hard decision boundaries, often struggle to capture such ambiguity. This paper introduces Fuzzy Ensemble Embedding Learning (FEEL), a novel fuzzy methodology that integrates fuzzy embeddings, entropy-based ensemble integration, and fuzzy-aware regularization into a unified framework. FEEL represents features through multi-membership fuzzy embeddings, dynamically weights classifiers using fuzzy entropy, and penalizes overconfident predictions in vague regions. Experimental validation is performed on benchmark datasets including UCI Vertigo, Breast Cancer, Credit Scoring, and MNIST with noisy labels. This study results demonstrate that FEEL consistently outperforms standard classifiers (SVM, Random Forest, Neural Networks), fuzzy rule-based systems, and fuzzy neural networks in terms of accuracy, robustness, and uncertainty calibration. The proposed method enhances interpretability through fuzzy membership visualization while maintaining scalability across domains. FEEL provides a general and flexible framework for real-world applications where ambiguity is unavoidable.

Highlights

• Introduces Fuzzy Ensemble Embedding Learning (FEEL) for scalable fuzzy machine learning.


• Embeds fuzzy membership functions at the feature level for classifier-agnostic integration.


• Incorporates entropy-based ensemble weighting to enhance robustness under uncertainty.


• Proposes fuzzy regularization to reduce overconfident predictions in ambiguous regions.


• Outlines empirical validation framework with medical science for dataset-specific results.