Smooth Receiver Operating Characteristics (smROC) Curves
Supervised learning algorithms perform common tasks including classification, ranking, scoring, and probability estimation. We investigate how scoring information, often produced by these models, is utilized by an evaluation measure. The ROC curve represents a visualization of the ranking performance of classifiers. However, they ignore the scores which can be quite informative. While this ignored information is less precise than that given by probabilities, it is much more detailed than that conveyed by ranking. This paper presents a novel method to weight the ROC curve by these scores. We call it the Smooth ROC (smROC) curve, and we demonstrate how it can be used to visualize the performance of learning models. We report experimental results to show that the smROC is appropriate for measuring performance similarities and differences between learning models, and is more sensitive to performance characteristics than the standard ROC curve.
KeywordsReceiver Operating Characteristic Receiver Operating Characteristic Curve Class Membership Brier Score Supervise Learning Algorithm
- 1.Asuncion, A., Newman, D.J.: UCI Machine Learning Repository (2007), http://www.ics.uci.edu/~mlearn/MLRepository.html
- 2.Bennett, P.N.: Using Asymmetric Distributions to Improve Text Classifier Probability Estimates. In: Proceedings of ACM SIGIR 2003, pp. 111–118 (2003)Google Scholar
- 7.Fawcett, T.: ROC Graphs: Notes and Practical Considerations for Data Mining Researchers. Technical Report HPL-2003-4, HP Labs (2003)Google Scholar
- 10.Grossman, D., Domingos, P.: Learning Bayesian Network Classifiers by Maximizing Conditional Likelihood. In: Proceedings of ICML 2004, pp. 361–368 (2004)Google Scholar
- 11.Ling, C.X., Huang, J., Zhang, H.: AUC: A Better Measure than Accuracy in Comparing Learning Algorithms. In: Proceedings of Canadian AI 2003, pp. 329–341 (2003)Google Scholar