Ensemble Feature Selection With Block-Regularized m × 2 Cross-Validation.

Area of Science:

• Machine Learning
• Bioinformatics
• Statistical Modeling

Background:

• Ensemble feature selection (EFS) is crucial for reducing noise and stabilizing results.
• Block-regularized m × 2 cross-validation shows superior performance in generalization and algorithm comparison.
• Traditional EFS often uses binomial distributions, which may not accurately represent feature selection frequencies.

Purpose of the Study:

• To propose a novel EFS technique utilizing block-regularized m × 2 cross-validation.
• To enhance the accuracy of feature selection by employing a more suitable distribution model.
• To theoretically and empirically demonstrate the advantages of the proposed EFS method.

Main Methods:

• Developed a new ensemble feature selection approach.
• Applied block-regularized m × 2 cross-validation for enhanced performance.
• Modeled feature selection frequency using a beta distribution, outperforming the traditional binomial distribution.

Main Results:

• The proposed technique offers higher selection probability for important features.
• It demonstrates a lower selected risk for noise features, leading to more true positives and fewer false positives.
• Theoretical analysis and experimental validation confirmed the effectiveness on simulated and real data.

Conclusions:

• The novel EFS method based on block-regularized m × 2 cross-validation is effective.
• Using a beta distribution for feature selection frequency improves accuracy and reliability.
• This approach offers significant advantages in identifying relevant features and minimizing noise in data analysis.