HeFS: helper-enhanced feature selection via Pareto-optimized genetic search.

Area of Science:

• Machine Learning
• Data Science
• Computational Biology

Background:

• Feature selection is vital for high-dimensional data but NP-hard.
• Conventional methods often yield suboptimal feature subsets due to premature convergence.
• Complex feature interactions in high-dimensional data pose challenges for existing techniques.

Purpose of the Study:

• To enhance feature selection by identifying complementary helper features.
• To develop a plug-and-play refinement framework for feature selection.
• To overcome limitations of heuristic and greedy approaches in high-dimensional settings.

Main Methods:

• Conditional Feature Selection formulated to augment existing subsets with complementary features from the residual space.
• Helper-Enhanced Feature Selection (HeFS) as a post-selection refinement framework compatible with various algorithms.
• Utilized biased initialization, ratio-guided mutation, and multi-objective search balancing accuracy and complementarity.

Main Results:

• HeFS consistently improved baseline feature selection performance across 18 benchmark datasets.
• Helper features boosted classification accuracy by over 9% compared to traditional methods and 6% over state-of-the-art.
• Helper features showed significantly lower correlation with baseline features, indicating reduced redundancy.

Conclusions:

• HeFS effectively models feature complementarity, addressing limitations of conventional methods.
• The framework systematically identifies informative features previously overlooked.
• Improved classification performance was observed across diverse application domains.