EMaTO-LFS: Evolutionary Many-Task Optimization-Based Localized Feature Selection for High-Dimensional Classification

Area of Science:

• Machine Learning
• Data Mining
• Computational Intelligence

Background:

• Localized feature selection (LFS) partitions data but often ignores inter-region correlations, limiting performance in high-dimensional (HD) scenarios.
• Existing LFS methods treat local regions independently, failing to leverage shared information crucial for complex datasets.

Purpose of the Study:

• To propose a novel evolutionary many-task optimization-based LFS (EMaTO-LFS) framework to address limitations of existing LFS methods in HD scenarios.
• To model LFS as a many-task optimization problem, considering regional correlations and enabling knowledge sharing among neighboring regions.

Main Methods:

• Developed an EMaTO-LFS framework that treats each local region as a distinct multiobjective feature selection task.
• Utilized sample ratios to adaptively construct local regions and employed a filter-based prefiltering method for feature subsets.
• Designed a subset-based mutation operator and a knowledge transfer strategy based on neighborhood relationships to avoid negative transfer.

Main Results:

• EMaTO-LFS achieved competitive balanced accuracy on 14 HD datasets.
• The method produced smaller feature subsets compared to state-of-the-art feature selection (FS) and LFS methods.
• Demonstrated effective knowledge sharing and collaborative solving of HD FS tasks.

Conclusions:

• The proposed EMaTO-LFS framework offers a significant advancement in localized feature selection for high-dimensional data.
• Collaborative solving and knowledge transfer among correlated regions enhance classification performance and reduce feature subset size.
• This approach effectively addresses the limitations of traditional LFS methods by integrating regional correlations into the optimization process.