Multitask Quantile Regression under the Transnormal Model.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Estimating multi-task quantile regression in high-dimensional settings presents challenges due to nonlinearity and nonnormality.
• Existing methods may struggle with robustness and efficiency when dealing with complex data distributions.

Purpose of the Study:

• To develop a robust and efficient method for high-dimensional multi-task quantile regression under the transnormal model.
• To address nonlinearity and nonnormality simultaneously using rank-based covariance regularization.

Main Methods:

• Proposed rank-based ℓ₁ penalization with positive definite constraints for sparse covariance matrix estimation.
• Introduced rank-based banded Cholesky decomposition regularization for banded precision matrix estimation.
• Utilized alternating direction method of multipliers and nearest correlation matrix projection.

Main Results:

• Derived a simple closed-form solution for multi-task quantile regression.
• Achieved an 'oracle'-like convergence rate, demonstrating high efficiency.
• Provided provable prediction intervals under high-dimensional settings.

Conclusions:

• The proposed rank-based method effectively combines quantile regression and covariance regularization for high-dimensional data.
• The method offers a good balance between robustness and efficiency, outperforming existing approaches.
• Demonstrated practical utility in analyzing protein mass spectroscopy data.