Brier Curves: a New Cost-Based Visualisation of Classifier Performance

             José Hernandez-Orallo, Peter Flach and Cèsar Ferri

• A curve to graphically understand and assess classifiers
• We assume that the classifier scores are posterior class probabilities
• This provides a natural way of choosing the thresholds
• This new curve depends on the quality of the probability estimates
• It shows the performance for the full range of operating conditions
• We can choose and discard classifiers depending on the operating conditions
• We can also combine classifiers in order to obtain a lower overall loss
• The area under the Brier curve for cost proportions is equal to the Brier score
• Complete details in the ICML'11 paper.
• An R script for drawing Brier cuves.