EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Predicting Customer Churn in Telecommunications: A Controlled Comparison of Machine Learning Models Using Mutual Information Feature Selection

Ernest Edem Dzisah

Interdisciplinary Journal of the African Alliance for Research, Advocacy and Innovation · 01.11.2026 · 2 (4)

DOI: 10.64261/hfpwc474 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Interdisciplinary Journal of the African Alliance for Research, Advocacy and Innovation, online erschienen am 01.11.2026. Verfasst von Ernest Edem Dzisah. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Customer churn significantly impacts telecommunications revenue and acquisition costs, but machine learning can proactively identify at-risk customers using historical data. This study compares four supervised algorithms for churn prediction: Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting. Using a dataset of 7,032 customer records, the study applied a stratified 80:20 train-test split. Preprocessing included median imputation and standardization for numerical features, and most-frequent imputation with one-hot encoding for categorical data. Mutual Information feature selection, using a fixed random state of 42, retained the ten highest-ranked features. All models shared this identical pipeline to ensure a controlled comparison. Performance was evaluated on an independent holdout test set using accuracy, precision, recall, F1-score, ROC AUC, and confusion matrices, alongside five-fold stratified cross-validation on the training data. On the holdout set, Logistic Regression achieved the highest accuracy (78.82%), while Gradient Boosting recorded 78.61% accuracy and the highest ROC AUC (83.60%), slightly outperforming Logistic Regression (83.39%). During cross-validation, Gradient Boosting maintained a superior mean ROC AUC of 84.48% ± 0.65%, compared to 84.29% ± 0.52% for Logistic Regression. However, Gradient Boosting identified only 51.60% of actual churn cases in the holdout set, underscoring the ongoing difficulty of accurately capturing leaving customers. Ultimately, the findings demonstrate that predictive performance can be closely matched across different algorithms. Reliable churn prediction requires a holistic approach using multiple evaluation metrics rather than relying on accuracy alone.

Zitieren

APA

Dzisah, E. E. (2026). Predicting Customer Churn in Telecommunications: A Controlled Comparison of Machine Learning Models Using Mutual Information Feature Selection. Interdisciplinary Journal of the African Alliance for Research, Advocacy and Innovation, 2(4). https://doi.org/10.64261/hfpwc474

Vancouver

Dzisah EE. Predicting Customer Churn in Telecommunications: A Controlled Comparison of Machine Learning Models Using Mutual Information Feature Selection. Interdisciplinary Journal of the African Alliance for Research, Advocacy and Innovation. 2026;2(4). doi:10.64261/hfpwc474

BibTeX
@article{Dzisah2026,
  author = {Dzisah, Ernest Edem},
  title = {Predicting Customer Churn in Telecommunications: A Controlled Comparison of Machine Learning Models Using Mutual Information Feature Selection},
  journal = {Interdisciplinary Journal of the African Alliance for Research, Advocacy and Innovation},
  year = {2026},
  volume = {2},
  number = {4},
  doi = {10.64261/hfpwc474}
}
Kostenlos starten →