Feature Selection for Anomaly Detection in Call Center Data

In this study, we present the process of designing machine learning models for the detection of call center agent malpractices. Based on the features extracted from audio recordings of a given telephone conversation, appropriate one-class support vector machine, isolation forest, and multivariate Gaussian models are trained, evaluated and compared in order to determine the best use case. The labeled data used in the experiments was obtained from a real call center and the results obtained indicate that the system is usable in a real-world scenario. The accuracy of used machine learning models are validated by using the F1 score as a metric.