Anomaly Pattern Analysis Based on Machine Learning on Real Telecommunication Data

Fraud and anomalies are serious problems in the telecommunication world. These problems can be detected with the machine learning based tools. In this study, we used a Call Detail Record (CDR) data that contains 417000 call records made to 217 different countries. Firstly, we performed call traffic analysis and use K-means clustering to segment multi-valued categorical variables. Secondly, we worked on supervised models such as XGBoost, Extra Trees, Random Forest, unsupervised model such as Isolation Forest, and a new Mixture of Experts model computing average prediction probabilities of supervised models for anomaly detection. Thirdly, we generated anomaly scores by using the sum of the predictions by five models and labeled the calls based on the anomaly score. Finally, we found that 1% of the calls were suspected of fraud and the results were in line with industry reports.