Towards Class-Imbalance Aware Multi-Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-label learning assigns multiple class labels to single instances.
• Class imbalance, where positive labels are rare, significantly degrades multi-label model performance.
• Existing methods often struggle to balance label correlation exploitation with imbalance handling.

Purpose of the Study:

• To propose a novel, effective, and class-imbalance aware learning strategy for multi-label classification.
• To simultaneously leverage label correlations and address class imbalance in multi-label datasets.
• To enhance the generalization performance of multi-label predictive models.

Main Methods:

• Introduced Cross-Coupling Aggregation (COCOA), a strategy designed to tackle class imbalance in multi-label learning.
• COCOA induces multiclass imbalance learners for each label by random coupling with other labels.
• Predictions from these learners are aggregated to determine labeling relevancy for unseen instances.

Main Results:

• COCOA demonstrated effectiveness in handling class-imbalance in multi-label learning.
• Extensive experiments on 18 benchmark datasets validated COCOA's superiority over state-of-the-art methods.
• Performance improvements were particularly notable in imbalance-specific evaluation metrics.

Conclusions:

• COCOA is a simple yet effective approach for multi-label learning with imbalanced data.
• The strategy successfully balances the exploitation of label correlations and the mitigation of class imbalance.
• COCOA offers a promising solution for improving multi-label classification accuracy in real-world scenarios.