Abstract:
Mobile network operators in developing countries often rely on traditional fraud detection systems, overlooking the
potential of advanced machine learning techniques. This study addresses this gap by developing an International Revenue
Share Fraud (IRSF) detection model using ensemble learning with random forest and support vector machine algorithms.
The model analyzes Call Detail Records (CDRs) to identify fraudulent call patterns. CDRs contain call attributes like
time, duration, source and destination numbers and completion status, providing valuable data for anomaly
detection. Random Forest is chosen for its effectiveness in handling complex and imbalanced datasets, common in
telecom fraud scenarios. Its ability to address imbalanced data is crucial, as fraudulent calls are typically rare compared
to legitimate ones. This research aims to develop a machine learning model that leverages call logs to detect fraudulent
international account takeover. Our results advance descriptive analysis and improve knowledge of the traits and patterns
of IRSFs. In the end, this produces a picture of IRSF operations that is more accurate. The model demonstrates good
predictive performance on the testing set with a Mean Absolute Error (MAE) of 1.1208, indicating a low average absolute
difference between predicted and actual values and the R-squared value of 0.9828 signifying strong overall predictive
accuracy
