Regularization in skewed binary classification

摘要: Skewed binary classification concerns the assignment of a new unknown object to one two populations, 0 or 1, on basis q-dimensional vector x = (x1, …xq), where for example population 0, is prevalent class. Assignment rules are developed from learning samples known objects, that is, objects come each populations. Since 1 rare class, overfitting and generalization problems arise easily many models. We propose an effective solution by assigning more weights class 1. The idea produce noisy replicates cases while keeping dominant unchanged. models considered are: nearest neighbor method, neural networks, trees, quadratic discriminant. Noisy replication was applied three real world simulated data sets. Encouraging results were obtained all considered.