Sparse Bayesian Broad Learning System via Adaptive Lasso Priors for Robust Regression.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computational Statistics

Background:

• Broad Learning System (BLS) excels in regression but is sensitive to outliers and noisy data.
• Existing BLS methods often use least squares, leading to reduced prediction accuracy with contaminated data.

Purpose of the Study:

• To introduce a robust sparse Bayesian Broad Learning System (BLS) using adaptive Lasso priors (AL-SBBLS).
• To enhance the performance of BLS in regression tasks affected by outliers and noise.

Main Methods:

• Applied adaptive Lasso constraints to enhance output weight sparsity and feature selection.
• Developed a multilayer Bayesian framework with adaptive Lasso priors for regularization and probability estimation.
• Utilized alternating direction method of multipliers (ADMMs) and variational Bayesian inference for network training.

Main Results:

• The proposed AL-SBBLS demonstrated superior robustness and predictive accuracy on 14 real-world datasets.
• Achieved the lowest average rank (1.44) in Friedman tests against 11 state-of-the-art BLS variants.
• Effectively mitigated the impact of outliers and noise through feature selection and probability distribution estimation.

Conclusions:

• AL-SBBLS offers significant improvements in robustness and accuracy for regression tasks with noisy data.
• The method successfully addresses the limitations of traditional BLS approaches in handling data contamination.
• AL-SBBLS represents a state-of-the-art advancement in robust machine learning for regression.