Multi-Label Feature Selection with Feature-Label Subgraph Association and Graph Representation Learning.

Area of Science:

• Machine Learning
• Data Mining
• Computational Science

Background:

• Multi-label data presents computational challenges due to high dimensionality and complex label dependencies.
• Effective feature selection is crucial for improving the performance of multi-label learning algorithms.
• Existing methods struggle with the intricate relationships between features and multiple labels.

Purpose of the Study:

• To propose a novel multi-label feature selection method named SAGRL.
• To effectively represent and leverage the complex correlations between features and labels.
• To enhance the accuracy and efficiency of feature selection in multi-label classification tasks.

Main Methods:

• Developed a graph representation learning approach (SAGRL) for multi-label feature selection.
• Mapped features and labels to nodes, establishing connections to form feature and label sets.
• Constructed feature-label subgraphs to capture abundant feature combinations and adjusted relationships via graph representation learning.

Main Results:

• The proposed SAGRL method demonstrated superior performance across six evaluation metrics.
• Experimental results on 11 datasets confirmed the effectiveness of the method.
• Achieved superior performance compared to several state-of-the-art multi-label feature selection techniques.

Conclusions:

• SAGRL effectively addresses the challenges of feature selection in multi-label data.
• The graph-based approach captures intricate feature-label relationships for improved selection.
• The method offers a promising solution for enhancing multi-label learning performance.