Coverage-performance curves for classification in datasets with a typical data

Handling atypical examples in classification tasks is one of the challenges in machine learning. While there seems to be a race for accuracy, very little has been done to understand and solve the issues related to atypical data. In this paper, coverage-performance (CP) curves are introduced to help a better understanding of atypical data. The concept of CP curves is based on the idea of separating atypical data and visualizing performance of classification as a function of coverage (the fraction of data participating in training or evaluation). To generate CP curves, two schemes are compared in this paper. The first scheme is based on SVMs alone and the second one is a hybrid of a PNN and a SVM. Two generated datasets with overlapping features are used to demonstrate the effectiveness of CP curves obtained by each scheme. Calculated theoretical limits on the generated data show that the hybrid scheme is a very effective way of producing CP curves. It is also shown that by separating atypical data, although we lose some data, the performance of the classification increases significantly.