To What Extent Can Machine Learning Detect Anomalous Call Behavior Using Telecommunications Metadata?

A very quick transition to digital services in Uzbekistan has increased the risk of cyber-enabled fraud. The ones to rise are scam calls that rely on social engineering techniques. Despite the increasing incidence of such cases, there is a limited amount of research that has truly studied technological approaches to detecting fraudulent calls in the Uzbek telecommunications environment. The following study investigates whether machine learning techniques can actually identify anomalous calling behavior using telecommunications metadata.

The analysis uses a dataset of 6,575,933 anonymized call records, from which a sample of 1,315,187 observations was selected. It is not a secret that there is a lack of labeled fraud indicators. Thus, the study applies an unsupervised anomaly detection approach using the Isolation Forest algorithm. The following algorithm features derived from temporal and event-based call attributes. The model identified 123,263 anomalous calls (1.87% of the dataset). Truly, these anomalies cannot be “claimed” to be called fraudulent calls. However, the findings illustrate that telecommunications metadata can reveal suspicious behavioral patterns and may support the development of AI-assisted telecom fraud detection systems.