Sci Rep. 2026 May 17. doi: 10.1038/s41598-026-53357-y. Online ahead of print.
ABSTRACT
This research presents a HashCode-based security abstraction that implements privacy-by-design in customer churn statistics, safeguarding identity while maintaining analytical integrity. A set of machine learning and deep learning models, including Logistic Regression, Random Forest, XGBoost, and a Multilayer Perceptron (MLP), is ideal for assessing churn while maintaining stringent privacy standards. Experimental results show that models that merely use behavioral, transactional, and temporal aspects can participate with and balance each other. The SGD Logistic Regression has an accuracy of 77.5%, a precision of 0.759, a recall of 0.812, an F1-score of 0.785, and an AUC of 0.815. This shows that it is quite sensitive to churners. The Random Forest has an accuracy of 77.0%, a precision of 0.767, a recall of 0.782, an F1-score of 0.775, and an AUC of 0.794. This means that it can generalize well. XGBoost has an accuracy of 74.0%, a precision of 0.733, a recall of 0.762, an F1-score of 0.748, and an AUC of 0.769. This shows that aggressive boosting doesn’t work very well with this structured dataset. The MLP (Keras) has the best overall performance, with an accuracy of 80.0%, a precision of 0.802, a recall of 0.802, an F1-score of 0.802, and the highest AUC of 0.825. This shows that it is better at learning non-linear representations and balancing categorization. These results provide strong churn prediction without relying on rich identifiers or assumptions that need a lot of infrastructure, which is different from previous studies. The study offers a deployable, regulation-compliant architecture demonstrating that accurate churn prediction is attainable by behavior-driven analytics within a consistent security enforcement paradigm.
PMID:42144427 | DOI:10.1038/s41598-026-53357-y
