Differential privacy-based evaporative cooling feature selection and classification with relief-F and random forests

Area of Science:

• Bioinformatics
• Machine Learning
• Statistical Learning

Background:

• Accurate classification from high-dimensional biological data is crucial but challenging.
• Feature selection improves accuracy but risks overfitting, especially in p >> n scenarios.
• Existing differential privacy methods struggle with overfitting in bioinformatics.

Purpose of the Study:

• Introduce a novel privacy-preserving machine learning algorithm for bioinformatics.
• Address overfitting in feature selection and classification with high-dimensional data.
• Develop a method inspired by statistical physics for robust privacy preservation.

Main Methods:

• Private Evaporative Cooling: a stochastic algorithm using Relief-F for feature selection and random forest for classification.
• Relates privacy threshold to thermodynamic Maxwell-Boltzmann distribution (temperature).
• Applies Evaporative Cooling concept from atomic gases for backward stepwise feature selection.

Main Results:

• Private Evaporative Cooling achieves higher classification accuracy without overfitting on simulated data with interactions.
• Comparable accuracies to thresholdout with random forest in simulations without interactions.
• Successfully applied to human brain resting-state fMRI data for major depressive disorder research.

Conclusions:

• Private Evaporative Cooling offers improved accuracy and overfitting prevention in privacy-preserving bioinformatics.
• The algorithm demonstrates effectiveness on complex, real-world biological datasets.
• Provides a robust approach for sensitive biological data analysis.