Detection is usually carried out following the Neyman-Pearson criterion to maximize the probability of detection (true positives rate), maintaining the probability of false alarm (false positives rate) below a given threshold. When the classes are unbalanced, the performance cannot be measured just in terms of true positives and false positives rates, and new metrics must be introduced, such as Precision. "Anger detection" in Interactive Voice Response (IVR) systems is one application where precision is important. In this paper, a cost function for features selection to maximize precision in anger detection applications is presented. The method has been proved with a real database obtained by recording calls managed by an IVR system, demonstrating its suitability.
展开▼
