Agent-Based Behavioural Modelling for Anomaly Detection in Call Data from Telecommunication Networks

Anomaly detection in telecommunications data tries to discover deviant behaviour of individual subscribers, including for example: detection of inconsistencies in call data, such as customer churn or attrition, potential fraud, deliberate or unintended expensive mistakes in call data, and so on. These have consequently led to unquantifiable loss of revenue to many telecommunication networks world-wide. Although the intentions of most subscribers to these networks are unknown when making phone calls, their service consumption pattern is reflected in their call data. Recent studies have investigated the challenges of anomaly detection but have not given conclusive solutions to address this problem holistically. In this work, we infer that if appropriate anomaly indicators for individual subscribers are used for detection, the true positive rates of the approaches can be maximized, while the false alarms can be minimized. The challenge addressed in this research is to find a technique of efficiently identifying the indicators that facilitate anomaly detection methods, providing qualitative knowledge. This knowledge will assist analysts and managers in explanatory, exploratory and decision making activities. We address this challenge by using an Agent-Based Behavioural Bayesian Network (BBN). This BBN models an individual subscriber by using an evolutionary algorithm, and anomaly indicators are identified through emergent behaviour. Hence, our implementation results for land-line subscribers provide a general approach of improving the detection qualities of existing anomaly detection methods in telecommunication networks.