Fast Multilabel Feature Selection via Global Relevance and Redundancy Optimization.

Area of Science:

• Machine Learning
• Data Mining

Background:

• Information-theoretic methods show promise for multilabel feature selection (MLFS).
• Existing MLFS methods often use heuristics, potentially leading to suboptimal results and high computational costs.
• Current approaches may not fully leverage labeling information.

Purpose of the Study:

• To propose a general optimization framework, Global Relevance and Redundancy Optimization (GRRO), for efficient and effective MLFS.
• To address limitations of existing methods in utilizing labeling information and computational efficiency.
• To develop an enhanced version, GRROfast, for faster MLFS by filtering inessential labels and features.

Main Methods:

• Developed the GRRO framework incorporating feature relevance, label relevance (correlation), and feature redundancy.
• Formulated MLFS to avoid repetitive entropy calculations for global optimality.
• Extended GRRO to GRROfast, featuring label/feature filtering and ensemble reconstruction for efficiency.
• Incorporated label-specific feature formulations to exploit multilabel data properties.

Main Results:

• GRRO achieves global optimal solutions efficiently by integrating relevance and redundancy.
• GRROfast significantly improves efficiency by filtering non-essential labels and features.
• The proposed algorithms effectively exploit multilabel data characteristics.
• Extensive experiments validate the effectiveness and efficiency of GRRO and GRROfast.

Conclusions:

• GRRO provides a robust and efficient framework for multilabel feature selection.
• GRROfast offers a computationally efficient extension for large-scale MLFS tasks.
• The developed methods enhance the utilization of label information and data properties in MLFS.