A high‐dimensional classification rule using sample covariance matrix equipped with adjusted estimated eigenvalues
Seungchul Baek, Hoyoung Park, Junyong Park · Stat · 2021
High‐dimensional classification has challenges mainly due to the singularity issue of the sample covariance matrix. In this work, we propose a different approach to get a more reliable sample covariance matrix by adjusting the estimated eigenvalues. This procedure also brings us a nonsingular matrix as a by‐product. We improve the optimization procedure to obtain a linear classifier by incorporating the adjusted sample covariance matrix and a shrinkage mean vector into the original optimization problem. We have shown that our proposed binary classification rule is better than some other rules in terms of the misclassification rate through most of various synthetic data and real data sets.