Subspace-Based Method to Improve Classification Accuracy of High-Dimensional Data