Benefiting From Single-Objective Feature Selection to Multiobjective Feature Selection: A Multiform Approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Evolutionary multiobjective feature selection (FS) faces challenges like duplicated solutions and low search efficiency.
• Traditional single-objective FS methods often use weighted sums, limiting exploration to specific directions.

Purpose of the Study:

• To propose a multiform framework for evolutionary multiobjective FS within a multitask environment.
• To enhance the diversity and efficiency of evolutionary algorithms for FS.

Main Methods:

• A novel multiform framework solving a multiobjective FS task alongside auxiliary single-objective FS tasks.
• Utilizing different direction vectors in auxiliary tasks to guide the evolutionary search.
• Comparative analysis against classical and state-of-the-art multiobjective evolutionary and FS algorithms.

Main Results:

• The proposed method demonstrates effectiveness and efficiency across 18 classification datasets.
• Verified performance improvements in both standard and noisy environments.
• The framework successfully boosts evolutionary search for multiobjective FS.

Conclusions:

• The proposed multitask framework effectively addresses diversity loss and improves search efficiency in evolutionary multiobjective FS.
• The integration of auxiliary single-objective tasks provides a promising direction for future FS research.