A Multi-Layered Machine Learning Framework for High-Volume Telecommunication Fraud Detection

Abstract

Telecommunication networks are faced with the task of managing unprecedented large streams of Call Detail Records (CDRs). Classical static approaches systems for managing fraud appear increasingly overwhelmed by Sensitivity dynamic. The notion of the sensitivity refers to the scenario where recent method of isolating sophisticated structural fraud signals inadvertently results to generated wrong false positive rates especially among high-traffic, genuine subscribers. To approach this challenge, this we present a highly scalable, multi-layered machine learning framework which is validated on a real time production dataset of 16.4 million Call Detail Records (CDRs). By integrating a Non-Homogeneous Poisson Process (NHPP) to model temporal non-stationarity with Fisher’s Linear Discriminant Analysis (LDA) for geometric feature projection, the proposed framework separates cyclic behavioral intensity from genuinely malicious signatures. A dynamically weighted Artificial Neural Network (ANN) then performs the final classification, achieving a fraud recall of 97.5% while reducing the false-positive rate to an operationally significant 1.73%.