Risk and Error Matrix charts

EasyChair Preprint 1218

Abstract

Measuring classifier performance is important for Machine Learning applications. There are various graphical representations relevant for the measures, such ROC. In this paper, Risk and Error matrix charts were developed; their characteristics were studied especially when dealing with the measure of the prevalence and incident population, particularly when their base rate of binary classification are not consistently selected. In our application, prevalence data is used as training set and incident data is used as scoring set. The methods to construct the charts were illustrated in order to gain the insight of how these charts are used for measuring classifier performance for several applications include weighted problem, class imbalance and detection of rare case problem. When the random sampling is not used for model building nor validation, and if incident data is required for independent model evaluation, then several procedures for correcting sampling incidence data are required in order to remove or reduce any bias measurements.

Keyphrases: binary classification, error matrix charts, risk matrix charts

@booklet{EasyChair:1218,
  author    = {Hari Koesmarno},
  title     = {Risk and Error Matrix charts},
  howpublished = {EasyChair Preprint 1218},
  year      = {EasyChair, 2019}}