ARCHIVES
VOL. 1, ISSUE 1 (2025)
Enhanced predictive data mining algorithm for fraud detection and churn modelling in telecommunication systems
Authors
Iwok Odudu Abasi Michael
Abstract
The telecommunications industry operates in an
intensely competitive and data-rich environment, where two of the most critical
challenges are fraudulent activities and customer churn. These issues result in
billions of dollars in annual revenue loss, making their effective mitigation a
top priority for operational sustainability and profitability. Traditional data
mining and statistical approaches often struggle with the high-dimensional,
imbalanced, and complex nature of telecom data, leading to suboptimal detection
rates and high false positives. This paper proposes an Enhanced Predictive Data
Mining Algorithm (EPDMA) that integrates advanced feature engineering, a hybrid
sampling technique to address class imbalance, and a stacked ensemble learning
model to simultaneously and more accurately address both fraud detection and
churn prediction. The feature engineering phase leverages Recursive Feature
Elimination (RFE) and domain-specific feature creation, such as rolling window
statistics and behaviour change metrics. To handle the severe class imbalance
inherent in both problems, a combination of Synthetic Minority Over-sampling
Technique (SMOTE) and Edited Nearest Neighbours (ENN) is employed. The core
predictive model is a stacked ensemble comprising diverse base learners,
including Gradient Boosting Machines (XGBoost), Random Forest, and a deep
neural network, whose outputs are meta-learned by a logistic regression model
for final prediction. The proposed EPDMA was rigorously evaluated on
real-world, anonymized telecom datasets and benchmarked against
state-of-the-art individual classifiers like XGBoost, Support Vector Machines
(SVM), and traditional logistic regression. The results demonstrate a superior
performance, with the EPDMA achieving an average precision of 94.5% and a
recall of 89.2% for fraud detection, and an AUC-ROC score of 0.935 for churn
modelling, significantly outperforming all baseline models. The discussion
elucidates how the integration of these techniques overcomes the limitations of
singular models, and the conclusion highlights the practical implications of
deploying such a robust system for telecom operators, ultimately leading to
enhanced security, improved customer retention strategies, and substantial
financial savings.
Download
Pages:7-11
How to cite this article:
Iwok Odudu Abasi Michael "Enhanced predictive data mining algorithm for fraud detection and churn modelling in telecommunication systems". World Journal of Engineering and Technology, Vol 1, Issue 1, 2025, Pages 7-11
Download Author Certificate
Please enter the email address corresponding to this article submission to download your certificate.