Weighted Mean Squared Deviation Feature Screening for Binary Features.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Mining

Background:

• Ultrahigh dimensional data presents challenges for feature selection in binary classification.
• Existing methods like Chi-square and mutual information may not optimally handle binary features with probabilities near 0.5.

Purpose of the Study:

• To propose a novel model-free feature screening method for ultrahigh dimensional binary classification.
• To address limitations of existing methods in identifying informative binary features.

Main Methods:

• Introduced weighted mean squared deviation (WMSD) for feature screening.
• Theoretically investigated asymptotic properties under the assumption log p = o(n).
• Employed a Pearson correlation coefficient method for practical feature selection based on power-law distribution.

Main Results:

• WMSD demonstrates superior performance compared to Chi-square and mutual information, particularly for features with probabilities near 0.5.
• The method's theoretical properties were rigorously examined.
• Empirical results on Chinese text classification show effectiveness with a small selected feature dimension.

Conclusions:

• The proposed WMSD method is a powerful tool for feature screening in ultrahigh dimensional binary classification.
• WMSD offers advantages in identifying relevant binary features, enhancing classification model performance.
• The method is validated through theoretical analysis and practical application in text classification.