This paper presents a heterogeneous uncertainty sampling approach for supervised learning, where one classifier selects instances for training another. The method is motivated by applications where the classifier used to select instances is computationally expensive. The authors tested using a probabilistic classifier to select examples for training the C4.5 rule induction program. The uncertainty samples produced by this approach yielded classifiers with lower error rates than random samples ten times larger. Uncertainty sampling involves iteratively requesting class labels for instances whose classes are uncertain. The method alternates between presenting instances to an expert for labeling and selecting instances whose labels are still uncertain. The type of classifier used in uncertainty sampling must be cheap to build and use. The authors' approach requires estimating the certainty of classifications, which is not always available in induction systems. The paper discusses the challenges of uncertainty sampling, including the distortion of class frequencies in uncertainty samples. The authors propose a correction method that proved effective and robust in experiments on a large text categorization dataset. They also discuss the use of a loss ratio parameter to adjust for the relative cost of false positives and false negatives in the C4.5 algorithm. The authors tested their approach on a dataset of news article titles, finding that uncertainty sampling with a probabilistic classifier produced significantly more accurate classifiers than random sampling. They also found that using a loss ratio greater than 1 helped counterbalance the overrepresentation of positive instances in uncertainty samples. The paper concludes that heterogeneous uncertainty sampling can be effective in reducing the number of instances an expert needs to label, and that the approach is particularly useful for text categorization tasks where large amounts of unlabeled data are available. The authors suggest that future work should focus on improving the efficiency of uncertainty sampling and better understanding the effects of heterogeneity in the sampling process.This paper presents a heterogeneous uncertainty sampling approach for supervised learning, where one classifier selects instances for training another. The method is motivated by applications where the classifier used to select instances is computationally expensive. The authors tested using a probabilistic classifier to select examples for training the C4.5 rule induction program. The uncertainty samples produced by this approach yielded classifiers with lower error rates than random samples ten times larger. Uncertainty sampling involves iteratively requesting class labels for instances whose classes are uncertain. The method alternates between presenting instances to an expert for labeling and selecting instances whose labels are still uncertain. The type of classifier used in uncertainty sampling must be cheap to build and use. The authors' approach requires estimating the certainty of classifications, which is not always available in induction systems. The paper discusses the challenges of uncertainty sampling, including the distortion of class frequencies in uncertainty samples. The authors propose a correction method that proved effective and robust in experiments on a large text categorization dataset. They also discuss the use of a loss ratio parameter to adjust for the relative cost of false positives and false negatives in the C4.5 algorithm. The authors tested their approach on a dataset of news article titles, finding that uncertainty sampling with a probabilistic classifier produced significantly more accurate classifiers than random sampling. They also found that using a loss ratio greater than 1 helped counterbalance the overrepresentation of positive instances in uncertainty samples. The paper concludes that heterogeneous uncertainty sampling can be effective in reducing the number of instances an expert needs to label, and that the approach is particularly useful for text categorization tasks where large amounts of unlabeled data are available. The authors suggest that future work should focus on improving the efficiency of uncertainty sampling and better understanding the effects of heterogeneity in the sampling process.