A Unified Multi-Class Feature Selection Framework for Microarray Data.

Area of Science:

• Machine Learning
• Data Science
• Computational Neuroscience

Background:

• Simultaneous multi-class feature selection is crucial for identifying informative features across all classes.
• Existing Recursive Feature Elimination (RFE) methods, effective for binary tasks, face computational and performance challenges when extended to multi-class problems.
• There is a need for efficient and effective multi-class feature selection techniques.

Purpose of the Study:

• To propose a unified multi-class feature selection (UFS) framework specifically designed for randomization-based neural networks.
• To address the computational cost and potential performance degradation associated with extending binary RFE methods to multi-class scenarios.
• To introduce a novel feature ranking criterion based on neural network output weights.

Main Methods:

• The UFS framework employs a randomization-based neural network.
• A new multi-class feature ranking criterion is proposed, based on the heuristic that feature importance correlates with the magnitude of output weights.
• Features are ranked using the norm of output weights and recursively eliminated based on their scores.

Main Results:

• Extensive experiments were conducted on 15 real-world datasets.
• The proposed UFS framework demonstrated superior performance compared to state-of-the-art multi-class feature selection algorithms.
• The framework effectively handles multi-class feature selection challenges.

Conclusions:

• The unified multi-class feature selection (UFS) framework offers an effective solution for multi-class problems.
• The proposed output weight-based ranking criterion is a viable approach for assessing feature importance in neural networks.
• The UFS framework provides a computationally efficient and high-performing alternative to existing methods.