Cost-sensitive multi-label learning with positive and negative label pairwise correlations.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-label learning assigns multiple labels to instances simultaneously.
• Binary Relevance (BR) is a common method but struggles with class imbalance and ignores label correlations.
• Existing methods rarely utilize negative label correlations, crucial for real-world applications.

Purpose of the Study:

• To propose a novel Cost-sensitive multi-label learning model with Positive and Negative Label (CPNL) pairwise correlations.
• To extend Binary Relevance to address class imbalance and exploit both positive and negative label correlations.
• To enhance model capability by providing a kernel extension for complex relationships and efficient solvers.

Main Methods:

• Developed a Cost-sensitive multi-label learning model (CPNL) extending Binary Relevance.
• Incorporated positive and negative label pairwise correlations into the model.
• Utilized accelerated gradient methods (AGM) for efficient linear and kernel model optimization.
• Provided a kernel extension to capture complex input-output relationships.

Main Results:

• CPNL effectively handles class imbalance in large label spaces.
• The model successfully leverages both positive and negative label correlations.
• CPNL demonstrates competitive performance against state-of-the-art multi-label learning approaches.
• Efficient optimization was achieved using accelerated gradient methods.

Conclusions:

• The proposed CPNL model offers a robust solution for multi-label learning challenges.
• CPNL's ability to model label correlations, including negative ones, improves performance.
• The method provides a competitive and efficient alternative for multi-label classification tasks.