Multilabel Feature Selection: A Local Causal Structure Learning Approach.

Area of Science:

• Machine Learning
• Data Science
• Causal Inference

Background:

• Multilabel feature selection is crucial for high-dimensional data.
• Existing methods often fail to consider all relevant feature correlations simultaneously.
• There's a need for scalable and comprehensive multilabel feature selection techniques.

Purpose of the Study:

• To propose a novel algorithm, M2LC, for multilabel feature selection.
• To address the simultaneous exploration of feature-label, feature-feature, and label-label correlations.
• To develop a scalable method for high-dimensional multilabel learning.

Main Methods:

• Formulating multilabel feature selection as a local causal structure learning problem.
• Developing the M2LC algorithm incorporating two novel error-correction subroutines.
• Learning local causal structures for each class label to identify feature relationships.

Main Results:

• M2LC simultaneously considers three types of feature relationships (feature-label, feature-feature, label-label).
• The algorithm is scalable to high-dimensional datasets.
• M2LC effectively corrects false discoveries caused by label-label correlations.
• Causally informative features are selected, and feature sharing is visualized.

Conclusions:

• M2LC offers a robust approach to multilabel feature selection by leveraging causal inference.
• The method provides insights into causal mechanisms and feature relationships within the data.
• M2LC outperforms state-of-the-art algorithms in extensive experimental evaluations.