A modification of the Nearest Shrunken Centroid (PAM) classifier that replaces its soft thresholding of class centroids with hard thresholding or order thresholding, with the thresholding parameter chosen by a deep-search procedure. Because soft thresholding biases the retained estimates and keeps too many features, these alternatives yield much sparser (fewer-gene) models with equal or better high-dimensional classification accuracy, demonstrated on multi-class human-cancer microarray datasets and Monte Carlo studies.
Studies how the choice of thresholding rule affects the Nearest Shrunken Centroid classifier, popularly known as Prediction Analysis of Microarrays (PAM), for high-dimensional classification. PAM shr…