Partial Multilabel Learning Using Noise-Tolerant Broad Learning System With Label Enhancement and Dimensionality...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Partial multilabel learning (PML) deals with noisy supervision where instances have an overcomplete set of candidate labels, but only a subset is correct.
• Existing label enhancement methods struggle with noisy label spaces, and few PML methods address label ambiguity, feature redundancy, and model efficiency simultaneously.

Purpose of the Study:

• To propose a novel joint partial multilabel framework using broad learning systems (BLS-PML) to overcome limitations in existing PML methods.
• To enhance label trustworthiness, reduce feature space redundancy, and improve model efficiency for robust PML.

Main Methods:

• A novel label enhancement method reconstructs a trustworthy label space, mitigating bias from noisy labels.
• A confidence-based dimensionality reduction method creates a low-dimensional feature space, reducing redundancy.
• A noise-tolerant broad learning system (BLS) is developed with dimensionality reduction and trustworthy label layers for PML.

Main Results:

• The BLS-PML framework was evaluated on six real-world and seven synthetic datasets against eight state-of-the-art algorithms.
• The proposed method significantly outperformed baseline algorithms in approximately 80% of 144 experimental scenarios.
• Demonstrated robustness and effectiveness in handling partial multilabel tasks.

Conclusions:

• The novel BLS-PML framework effectively addresses key challenges in partial multilabel learning, including noisy labels and feature redundancy.
• The proposed mechanisms for label enhancement, dimensionality reduction, and noise tolerance lead to superior performance.
• The study highlights the potential of broad learning systems for complex machine learning tasks like PML.