Regularization in skewed binary classification (Q1966373)

The paper deals with the problem of overfitting in classification of a new unknown object to one of two populations, 0 and 1, on the basis of a q-dimensional explanatory vector \(x=(x_1,\dots,x_q)\), where one of the populations, for example population 0, is the prelevent class and population 1 is a rare class. The author proposed a solution to this problem by increasing the occurrence of the rare cases by producing noisy replicates in the training data from the rare cases while keeping the objects from the dominant class without changes. He studies the effect of adding noise during training for several classification approaches: nearest neighbor method, neural networks, classification trees and quadratic discriminants. Computer experiments on three data sets from the Information and Computer Science repository of the University of California at Urvine were carried out. Promising and encouraging results are obtained.