Quarter-Sphere Support Vector Machine for Fraud Detection in Mobile Telecommunication Networks

This paper addresses the problem of finding out fraudulent calls of mobile phone users by comparing the most recent call patterns with their past usage patterns. We have modeled the user's profile based on the most relevant fraud detection features like call duration, call type, call frequency along with location and time data. The Reality Mining dataset has been used for testing the efficiency of the proposed methodology. In this work, we discriminate the malicious behavior of users from the normal behavior by training on Support Vector Machine (SVM) classifier. An anomaly is detected when the current pattern of a user (subject) does not match with any of the individual's normal patterns. We have also focused on the improvement of the classifier by applying the concept of Quarter-Sphere SVM. The Quarter Sphere-SVM is a formulation of One-Class SVM, supported by Support Vector Data Description which helps the SVM in unsupervised learning. Our experiments show promising results in terms of detecting fraudulent calls without raising too many false alarms.