Enhanced Predictive Data Mining Algorithm for Fraud Detection and Churn Behaviour Modelling In Telecommunication Systems

Abstract

This work presents an enhanced predictive model for fraud detection and churn behaviour modelling in a telecommunication network. Computational analytic modelling was employed by using probabilistic models; Naïve Bayesian model, linear discriminant function and neural prediction networks to achieve adaptive control policy for fraud/churn detection. A critical threshold discriminant function (CTDF) Value of 0.00229 was obtained from a multivariate analysis of samples of call detail record (CDR) data sets. From the neural network training and validation plot, the proposed data mining predictive model gave 1.7562 Mean Square Error (MSE) for the CTDF. Also, an evaluation was carried out to determine an optimal algorithm/model with accurate, consistent and reliable results. Hence, three algorithms were analysed for Fraud and churn behavioural mining/detections, viz Decision Tree (DT), Logistic Regression (LR) and Enhanced Neural Discriminant Analysis (Proposed). These gave 14.29%, 30.00%, and 55.71% respectively. It was therefore concluded that the proposed algorithm offers the best and most reliable prediction threshold for churn/fraud attrition.