Department of Agronomy, N.T.U.

   In another study, we adopted the coefficient of intrinsic dependence (CID) to identify putative signatures for classification. Our results showed that CID is promising in supervised learning. The simulations had shown that CID is robust in selecting features with different means or different variances in two classes. When applying to the breast cancer clinical array data, the genes selected by CID best classified ER+/−patients. However, CID is not appropriate to be immediately applied to unsupervised learning. Although the misclassification rate of CID was as low as those of conventional methods in most of the cases, CID suffered the curse of dimensionality the most. The small sample size relative to the number of variables (genes) is of particular concern in the microarray studies. When the sample size of the training data is small, one might yield a classifier that perfectly classifies the training sample but performs badly in the other samples. While applying CID in classification, the curse of dimensionality strikes from another direction. One may not observe a particular set of data in the training set but the same data appears in the test set. In this scenario, the probability that the object belonging to certain group is not estimable. Another way to estimate the conditional distribution, such as nonparametric smoothing, might solve the problem.