Robust statistical boosting with quantile-based adaptive loss functions.

Area of Science:

• Biostatistics
• Machine Learning
• Bioinformatics

Background:

• High-dimensional biomedical data often contains outliers, complicating predictive modeling and variable selection.
• Traditional methods struggle with data corrupted by outliers, leading to inaccurate results.
• Robust statistical methods are needed to handle such data effectively.

Purpose of the Study:

• To develop and evaluate adaptive robust loss functions within statistical boosting algorithms.
• To enhance variable selection and predictive modeling for high-dimensional biomedical data with potential outliers.
• To improve robustness against vertical outliers in the outcome variable.

Main Methods:

• Proposed an adaptive approach for composite robust loss functions (Huber, Bisquare) in boosting.
• Adapted the threshold parameter of loss functions based on residual sizes in each boosting iteration.
• Compared performance against M-regression, standard boosting losses, and lasso using simulated data and NCI-60 cell line expression data.

Main Results:

• Adaptive Huber and Bisquare losses demonstrated superior performance in prediction accuracy and variable selection when data contained outliers or corruption.
• For non-corrupted data, the adaptive approach performed comparably to standard L2 loss boosting and lasso.
• In analyzing KRT19 protein expression data, the adaptive loss functions yielded favorable prediction accuracy and highly sparse models.

Conclusions:

• Adaptive robust loss functions integrated with boosting algorithms offer a powerful tool for analyzing complex biomedical data.
• The proposed method effectively handles outliers, leading to more reliable variable selection and predictive models.
• This approach shows promise for applications in bioinformatics, particularly with gene expression data.