Evolutionary Dimension-Specific Feature Selection for Multi-Dimensional Classification.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multi-dimensional classification (MDC) involves instances with multiple interdependent class labels.
• Current MDC methods often fail to utilize distinct feature representations for different semantic dimensions.
• Irrelevant and redundant features can negatively impact model performance in MDC.

Purpose of the Study:

• To propose a feature selection approach for MDC that addresses the need for dimension-specific feature representations.
• To improve the performance of MDC models by mitigating the impact of irrelevant and redundant features.
• To introduce a method that accounts for interdependencies between class dimensions during feature selection.

Main Methods:

• A feature selection approach named Fest, based on evolutionary multi-tasking, is proposed for MDC.
• Fest treats feature selection for each class dimension as an independent evolutionary subtask.
• An exploration mechanism for feature interaction is incorporated to handle correlated class dimensions and dependencies.

Main Results:

• Extensive experiments were conducted on eleven benchmark datasets and four state-of-the-art MDC approaches.
• The results demonstrate that selecting dimension-specific features significantly improves classification performance compared to using all features.
• Fest effectively captures the semantics of each dimension and handles feature interactions between correlated dimensions.

Conclusions:

• Dimension-specific feature selection is crucial for enhancing MDC performance.
• Fest provides an effective evolutionary multi-tasking framework for feature selection in MDC.
• The proposed approach offers a significant improvement over existing MDC methods by optimizing feature subsets for each dimension.