An apparent paradox: a classifier based on a partially classified sample may have smaller expected error rate than that if the sample were completely classified

Abstract

There has been increasing interest in using semi-supervised learning to form a classifier. As is well known, the (Fisher) information in an unclassified feature with unknown class label is less (considerably less for weakly separated classes) than that of a classified feature which has known class label. Hence in the case where the absence of class labels does not depend on the data, the expected error rate of a classifier formed from the classified and unclassified features in a partially classified sample is greater than that if the sample were completely classified. We propose to treat the labels of the unclassified features as missing data and to introduce a framework for their missingness as in the pioneering work of Rubin (Biometrika 63:581–592, 1976) for missingness in incomplete data analysis. An examination of several partially classified data sets in the literature suggests that the unclassified features are not occurring at random in the feature space, but rather tend to be concentrated in regions of relatively high entropy. It suggests that the missingness of the labels of the features can be modelled by representing the conditional probability of a missing label for a feature via the logistic model with covariate depending on the entropy of the feature or an appropriate proxy for it. We consider here the case of two normal classes with a common covariance matrix where for computational convenience the square of the discriminant function is used as the covariate in the logistic model in place of the negative log entropy. Rather paradoxically, we show that the classifier so formed from the partially classified sample may have smaller expected error rate than that if the sample were completely classified.