An evolutionary decomposition-based multi-objective feature selection for multi-label classification

Area of Science:

• Data Mining and Machine Learning
• Computational Intelligence

Background:

• Multi-label classification assigns multiple categories to data entities, crucial for real-world applications.
• Feature selection is vital for improving multi-label classification by reducing irrelevant and redundant features.
• Conflicting objectives in multi-label feature selection include minimizing features and maximizing classification accuracy.

Purpose of the Study:

• To develop a multi-objective optimization algorithm for effective multi-label feature selection.
• To address the challenge of balancing feature reduction with classification performance.

Main Methods:

• An enhanced decomposition-based multi-objective optimization algorithm was developed.
• The algorithm divides the multi-label feature selection into solvable single-objective subproblems using evolutionary algorithms.
• A local search operator and a pool of genetic operators were incorporated to enhance solution quality and diversity.

Main Results:

• The proposed algorithm demonstrated improved performance on eight benchmark datasets compared to existing multi-objective feature selection methods.
• Evaluation metrics like hypervolume indicator and set coverage showed superior results.
• The method achieved higher classification accuracy with fewer features.

Conclusions:

• The developed multi-objective optimization algorithm effectively addresses the conflicting goals of multi-label feature selection.
• The approach enhances optimization speed and finds diverse feature subsets.
• This method offers a promising solution for improving multi-label classification accuracy and efficiency.