Telecommunication Fraud Detection Using Data Mining techniques
João Vitor, Cepêda de Sousa · 2014
This document presents the final report of the thesis “Telecommunication Fraud Detection Using Data Mining Techniques”, were a study is made over the effect of the unbalanced data, generated by the Telecommunications Industry, in the construction and performance of classifiers that allows the detection and prevention of frauds. In this subject, an unbalanced data set is characterized by an uneven class distribution where the amount of fraudulent instances (positive) is substantially smaller than the amount normal instances (negative). This will result in a classifier which is most likely to classify data has belonging to the normal class then to the fraud class. At first, an overall inspection is made over the data characteristics and the Naive Bayes model, which is the classifier selected to do the anomaly detection on these experiments. After the characteristics are presented, a feature engineering stage is done with the intent to extend the information contained in the data creating a depper relation with the data itself and model characteristics. A previously proposed solution that consists on undersampling the most abundant class (normal) before building the model, is presented and tested. In the end, the new proposals are presented. The first proposal is to study the effects of changing the intrinsic class distribution parameter in the Naive Bayes model and evaluate its performance. The second proposal consists in estimating margin values that when applied to the model output, attempt to bring more positive instances from previous negative classification. All of these suggested models are validated over a monte-carlo experiment, using data with and without the engineered features.