Predicting and optimizing classifier utility with the power law

When data collection is costly and/or takes a significant amount of time, an early prediction of the classifier performance is extremely important for the design of the data mining process. Power law has been shown in the past to be a good predictor of decision-tree error rates as a function of the sample size. In this paper, we show that the optimal training set size for a given dataset can be computed from a learning curve characterized by a power law. Such a curve can be approximated using a small subset of potentially available data and then used to estimate the expected trade-off between the error rate and the amount of additional observations. The proposed approach to projected optimization of classifier utility is demonstrated and evaluated on several benchmark datasets.