The problem of determining a proper rejection threshold in classification applications, along with the necessity to define an evaluation criterion for the resulting classifier, are major topics in the process of development of real-world recognition systems. This report reviews the concepts of error, rejection and performance evaluation in the framework of statistical pattern recognition, focusing on the concept of Equal Error Rate (ERR), which is proposed as an objective, threshold-independent measure of the performance of a classifier. A case study is discussed, i.e., evaluation of a speaker identification system based on neural networks, both in terms of error rate and of Equal Error Rate
Rejection and the Equal Error Rate: Principles and a Case Study
Trentin, Edmondo
1995-01-01
Abstract
The problem of determining a proper rejection threshold in classification applications, along with the necessity to define an evaluation criterion for the resulting classifier, are major topics in the process of development of real-world recognition systems. This report reviews the concepts of error, rejection and performance evaluation in the framework of statistical pattern recognition, focusing on the concept of Equal Error Rate (ERR), which is proposed as an objective, threshold-independent measure of the performance of a classifier. A case study is discussed, i.e., evaluation of a speaker identification system based on neural networks, both in terms of error rate and of Equal Error RateI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
