Nonparametric IPSS: fast, flexible feature selection with false discovery control

Area of Science:

• Machine Learning
• Statistical Methods
• Bioinformatics

Background:

• Feature selection is crucial in machine learning and statistics.
• Existing methods often rely on parametric models, lack theoretical false discovery control, or identify limited true positives.

Purpose of the Study:

• Introduce a general, nonparametric feature selection method with finite-sample false discovery control.
• Enhance the identification of true positives while maintaining statistical rigor.
• Provide efficient and accurate tools for high-dimensional data analysis.

Main Methods:

• Utilize Integrated Path Stability Selection (IPSS) applied to arbitrary feature importance scores.
• Develop specific implementations: IPSS Gradient Boosting (IPSSGB) and IPSS Random Forests (IPSSRF).
• Estimate q-values for better suitability in high-dimensional settings compared to P-values.

Main Results:

• IPSSGB and IPSSRF demonstrate accurate false discovery rate control in nonlinear simulations.
• Both methods significantly outperform existing approaches in detecting true positives.
• Achieve high efficiency, running in under 20 seconds for datasets with 500 samples and 5000 features.
• Applied to cancer data, IPSSGB and IPSSRF yield improved predictions with fewer features.

Conclusions:

• IPSS offers a powerful and flexible framework for feature selection in high-dimensional data.
• The developed IPSSGB and IPSSRF methods provide accurate and efficient solutions for biological data analysis.
• These methods advance the field by improving both statistical control and discovery power in feature selection.