Classifying Loss Rates with Small Samples T. X Brown in Proc. of IWANNT2, ed. J. Alspector, et al., Erlbaum, 1995 Abstract: In a modern telecommunication systems, loss rates are difficult to compute and, with sample sizes small compared to the loss rate, are difficult to measure empirically. When, based on such small samples, a classifier tries to identify what conditions result in loss rates that exceed a threshold, the decision boundary can miss the threshold by orders of magnitude. We develop two correction methods. The first weights samples to remove the bias. The second calculates confidence levels and combines these with sample aggregation to produce confident decisions. We analyze the methods and confirm the results using synthetic data.