Many-Objective Jaccard-Based Evolutionary Feature Selection for High-Dimensional Imbalanced Data Classification.

Area of Science:

• Computational intelligence
• Machine learning
• Data mining

Background:

• Feature selection (FS) methods, filters and wrappers, face challenges in high-dimensional, many-objective, imbalanced datasets.
• Evolutionary wrapper-based FS shows promise but requires efficient handling of computational costs and performance metrics.

Purpose of the Study:

• To propose a novel Jaccard similarity (JS)-based evolutionary many-objective (JSEMO) feature selection approach.
• To concurrently address evolutionary FS and imbalanced classifier selection.
• To investigate the mutual influence between feature selection and classifier choice.

Main Methods:

• JSEMO integrates JS into population initialization, reproduction, and elitism for enhanced diversity and duplicate solution avoidance.
• A set-based variation operator using intersection and union operators is employed for binary coding compatibility.
• A double-weighted KNN (KNN2W) classifier with four objectives is introduced for imbalanced data handling.

Main Results:

• JSEMO generated distinct optimal features across 15 benchmark problems, outperforming 20 existing methods.
• Significant improvements were observed in overall accuracy, balance accuracy, and g-mean metrics.
• Comparable feature set size and computational cost were maintained.

Conclusions:

• JSEMO effectively addresses challenges in high-dimensional, many-objective, imbalanced feature selection.
• The integration of JS and the set-based variation operator positively impacts algorithm performance.
• KNN2W with appropriate metrics is crucial for handling imbalanced distributions in many-objective FS.