Modal-Regression-Based Broad Learning System for Robust Regression and Classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Broad learning system (BLS) demonstrates strong performance in regression and classification.
• Standard BLS models are sensitive to noise and outliers due to their reliance on the least-squares criterion.
• Performance degradation in BLS is a significant challenge when dealing with contaminated datasets.

Purpose of the Study:

• To propose a robust modal-regression-based broad learning system (MRBLS) for handling noisy and outlier-corrupted data.
• To enhance the accuracy and robustness of broad learning systems in regression and classification tasks.
• To introduce an efficient training method for the proposed MRBLS model.

Main Methods:

• Developed a modal-regression-based BLS (MRBLS) by replacing the minimum mean square error (MMSE) criterion with modal regression for output weight training.
• Incorporated an l_2,1-norm-induced constraint to promote row sparsity in the connection weight matrix, enabling effective feature selection.
• Utilized the half-quadratic theory for efficient and effective optimization of the MRBLS network.

Main Results:

• The proposed MRBLS method demonstrated superior performance on various regression and classification datasets.
• MRBLS showed significant improvements in both accuracy and robustness compared to existing state-of-the-art BLS methods.
• Experimental results validated the effectiveness of MRBLS in handling data corrupted by noise and outliers.

Conclusions:

• The modal-regression-based broad learning system (MRBLS) offers enhanced robustness against noisy and outlier-contaminated data.
• MRBLS provides a more reliable alternative to conventional BLS models for real-world applications with imperfect data.
• The proposed method effectively combines modal regression, l_2,1-norm regularization, and half-quadratic optimization for robust machine learning.