Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Intuitionistic Fuzzy Broad Learning System: Enhancing Robustness Against Noise and Outliers

M. Sajid, A. K. Malik, M. Tanveer

TL;DR

The paper tackles the vulnerability of Broad Learning System (BLS) to noise and outliers by introducing two robust variants, F-BLS and IF-BLS, which weight samples through fuzzy and intuitionistic fuzzy schemes, respectively. F-BLS applies fuzzy membership in the original feature space, while IF-BLS operates in kernel space using membership and non-membership via neighborhood information, leading to improved generalization. The authors provide closed-form optimization formulations with two solution regimes, analyze computational complexity, and validate the approaches through extensive UCI benchmarks, Gaussian-noise tests, and Alzheimer's disease diagnosis using the ADNI dataset, demonstrating superior robustness and accuracy over strong baselines. The work contributes a practical, scalable enhancement to BLS that better handles noise and outliers, with public code and promising implications for biomedical classification tasks like AD diagnosis.

Abstract

In the realm of data classification, broad learning system (BLS) has proven to be a potent tool that utilizes a layer-by-layer feed-forward neural network. However, the traditional BLS treats all samples as equally significant, which makes it less robust and less effective for real-world datasets with noises and outliers. To address this issue, we propose fuzzy broad learning system (F-BLS) and the intuitionistic fuzzy broad learning system (IF-BLS) models that confront challenges posed by the noise and outliers present in the dataset and enhance overall robustness. Employing a fuzzy membership technique, the proposed F-BLS model embeds sample neighborhood information based on the proximity of each class center within the inherent feature space of the BLS framework. Furthermore, the proposed IF-BLS model introduces intuitionistic fuzzy concepts encompassing membership, non-membership, and score value functions. IF-BLS strategically considers homogeneity and heterogeneity in sample neighborhoods in the kernel space. We evaluate the performance of proposed F-BLS and IF-BLS models on UCI benchmark datasets with and without Gaussian noise. As an application, we implement the proposed F-BLS and IF-BLS models to diagnose Alzheimer's disease (AD). Experimental findings and statistical analyses consistently highlight the superior generalization capabilities of the proposed F-BLS and IF-BLS models over baseline models across all scenarios. The proposed models offer a promising solution to enhance the BLS framework's ability to handle noise and outliers. The source code link of the proposed model is available at https://github.com/mtanveer1/IF-BLS.

Intuitionistic Fuzzy Broad Learning System: Enhancing Robustness Against Noise and Outliers

TL;DR

The paper tackles the vulnerability of Broad Learning System (BLS) to noise and outliers by introducing two robust variants, F-BLS and IF-BLS, which weight samples through fuzzy and intuitionistic fuzzy schemes, respectively. F-BLS applies fuzzy membership in the original feature space, while IF-BLS operates in kernel space using membership and non-membership via neighborhood information, leading to improved generalization. The authors provide closed-form optimization formulations with two solution regimes, analyze computational complexity, and validate the approaches through extensive UCI benchmarks, Gaussian-noise tests, and Alzheimer's disease diagnosis using the ADNI dataset, demonstrating superior robustness and accuracy over strong baselines. The work contributes a practical, scalable enhancement to BLS that better handles noise and outliers, with public code and promising implications for biomedical classification tasks like AD diagnosis.

Abstract

In the realm of data classification, broad learning system (BLS) has proven to be a potent tool that utilizes a layer-by-layer feed-forward neural network. However, the traditional BLS treats all samples as equally significant, which makes it less robust and less effective for real-world datasets with noises and outliers. To address this issue, we propose fuzzy broad learning system (F-BLS) and the intuitionistic fuzzy broad learning system (IF-BLS) models that confront challenges posed by the noise and outliers present in the dataset and enhance overall robustness. Employing a fuzzy membership technique, the proposed F-BLS model embeds sample neighborhood information based on the proximity of each class center within the inherent feature space of the BLS framework. Furthermore, the proposed IF-BLS model introduces intuitionistic fuzzy concepts encompassing membership, non-membership, and score value functions. IF-BLS strategically considers homogeneity and heterogeneity in sample neighborhoods in the kernel space. We evaluate the performance of proposed F-BLS and IF-BLS models on UCI benchmark datasets with and without Gaussian noise. As an application, we implement the proposed F-BLS and IF-BLS models to diagnose Alzheimer's disease (AD). Experimental findings and statistical analyses consistently highlight the superior generalization capabilities of the proposed F-BLS and IF-BLS models over baseline models across all scenarios. The proposed models offer a promising solution to enhance the BLS framework's ability to handle noise and outliers. The source code link of the proposed model is available at https://github.com/mtanveer1/IF-BLS.
Paper Structure (14 sections, 19 equations, 3 figures, 5 tables, 1 algorithm)

This paper contains 14 sections, 19 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Broad Learning System (BLS) chen2017broad
  4. Fuzzy Membership Scheme Lin2002
  5. Intuitionistic Fuzzy Membership (IFM) Scheme ha2013support
  6. Proposed Fuzzy Broad Learning System (F-BLS) and Intuitionistic Fuzzy Broad Learning System (IF-BLS)
  7. Complexity Analysis of the Proposed F-BLS and IF-BLS
  8. Experiments and Results Analysis
  9. Setup for Experiments
  10. Results and Statistical Analyses on UCI Dataset
  11. Evaluation on UCI Datasets with Gaussian Noise
  12. Evaluation on ADNI Dataset
  13. Sensitivity Analyses of the Hyperparameters of the Proposed Models
  14. Conclusion and Future Work

Figures (3)

  • Figure 1: The architecture of the BLS model.
  • Figure 2: The flowchart of finding the output weight matrix $W_T$ of the proposed F-BLS and IF-BLS models.
  • Figure 3: Accuracy along with the standard deviation (shown in thin lines on the top) with respect to regularization parameter $C$ on ADNI dataset.