Compact Broad Learning System Based on Fused Lasso and Smooth Lasso

Area of Science:

• Machine Learning
• Artificial Intelligence
• Network Science

Background:

• Broad Learning System (BLS) offers a simplified network structure.
• Node correlations are crucial for effective BLS modeling.
• Existing BLS methods may not fully exploit inter-node correlations.

Purpose of the Study:

• To propose a novel simplification method for BLS called compact BLS (CBLS).
• To enhance BLS network structure compactness while preserving node correlations.
• To achieve a sparse and smooth output weight solution reflecting group learning characteristics.

Main Methods:

• Introduced CBLS, incorporating L1-regularization and fusion terms.
• Utilized L1-regularization to penalize output weights, capturing node-output correlations.
• Employed fusion terms (inspired by Fused Lasso and Smooth Lasso) to penalize adjacent output weight differences, capturing node-node correlations.
• Optimized output weights iteratively to simultaneously consider both types of correlations.

Main Results:

• Achieved significant simplification of the BLS network structure.
• Provided a sparse and smooth output weight solution.
• Demonstrated that simplification did not reduce prediction accuracy.
• Validated the effectiveness and feasibility of CBLS through experiments on public datasets.

Conclusions:

• CBLS offers a more reasonable network simplification for BLS.
• The method effectively captures and utilizes node correlations.
• The resulting sparse and smooth weights reflect the group learning nature of BLS.
• The proposed approach enhances BLS efficiency and interpretability.