Sparse quadratic classification rules via linear dimension reduction.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• High-dimensional data presents challenges for traditional classification methods, especially with unequal covariance matrices.
• Estimating full quadratic discriminant rules can be computationally intensive and unstable in high dimensions.

Purpose of the Study:

• To develop a scalable and robust classification method for high-dimensional data with unequal covariance matrices.
• To perform simultaneous variable selection and linear dimension reduction prior to classification.

Main Methods:

• A novel framework combining variable selection and linear dimension reduction.
• Subsequent application of quadratic discriminant analysis in the reduced space.
• Theoretical guarantees on variable selection consistency and empirical comparisons.

Main Results:

• The proposed method scales linearly with the number of measurements, outperforming traditional methods on high-dimensional datasets.
• Demonstrated variable selection consistency.
• Successfully applied to breast cancer gene expression data.

Conclusions:

• The developed method offers a computationally efficient alternative to full quadratic discriminant analysis for high-dimensional data.
• The approach highlights the importance of the ESR1 gene in differentiating estrogen receptor status in breast cancer.