Accurate and Efficient Large-Scale Multi-Label Learning With Reduced Feature Broad Learning System Using Label...

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Large-scale multi-label learning presents significant challenges due to numerous labels and complex data structures.
• Existing methods often suffer from poor classification performance or excessive training times on massive datasets.
• Broad Learning System (BLS) offers a succinct structure suitable for large-scale tasks but requires adaptation for complex label spaces.

Purpose of the Study:

• To propose a novel multi-label classifier, BLS-MLL, designed for large-scale datasets.
• To enhance classification performance and training efficiency in multi-label learning scenarios.
• To address the limitations of existing BLS models in handling large and complex label spaces.

Main Methods:

• Introduced a kernel-based feature reduction module with three layers: feature mapping (using elastic network regularization), enhancement nodes (kernel method for nonlinear conversion), and feature reduction.
• Incorporated a correlation-based label thresholding mechanism for effective conversion of decision values to logical outputs.
• Leveraged the inherent structure of Broad Learning System for efficient processing of large-scale data.

Main Results:

• BLS-MLL demonstrated superior performance, outperforming six state-of-the-art multi-label classifiers in 86% of tested cases.
• The proposed method achieved better training efficiency in 90% of the experimental comparisons.
• The kernel-based feature reduction and label thresholding mechanisms significantly improved accuracy and efficiency on high-dimensional, noisy data.

Conclusions:

• BLS-MLL effectively addresses the grand challenge of large-scale multi-label learning.
• The novel mechanisms enhance both classification accuracy and computational efficiency.
• The proposed approach offers a promising solution for complex multi-label classification tasks.