An enhanced gravitational search algorithm for feature selection in telecom churn prediction

• An enhanced feature selection method (GRAS) is proposed for churn prediction. • This method combines Gravitational Search Algorithm with Simulated Annealing. • Benchmarked with KNN and RF on four public telecom datasets, it outperforms GA, GSA, and SA. • GRAS selects smaller feature sets and runs faster than GA/GSA; results are Friedman-Nemenyi validated. • SHAP analysis aligns with GRAS-selected features, supporting interpretability and scalability. Customer churn poses a significant challenge to telecommunications companies, as retaining existing customers is more cost-effective than acquiring new ones. Accurate churn prediction enables timely interventions, reduces revenue loss, and enhances customer satisfaction. We propose GRAS, a feature selection method that integrates the global exploration of the Gravitational Search Algorithm (GSA) with the local refinement of Simulated Annealing (SA), aiming to improve the efficiency and interpretability. We evaluate GRAS with two base learners, k -Nearest Neighbors (KNN) and Random Forest (RF), on four publicly available churn datasets and benchmark its performance against metaheuristic baselines: Genetic Algorithm (GA), standalone GSA, and SA. The largest dataset contains 58 features and 51,047 samples and is included to stress-test scalability. The other three are smaller, with 11–21 features and 3,333–7,043 samples. On the largest dataset, GRAS with KNN attains the best Accuracy, Precision, Recall, F1, and AUC among GA, GSA, and SA. With RF, GRAS remains competitive across datasets and consistently selects the smallest feature subsets (lowest OFS), yielding compact and interpretable models. GRAS is markedly faster than GSA and GA, though slower than single-trajectory SA. These differences are confirmed by the Friedman test with Nemenyi post-hoc analysis. To support transparency, we conduct a SHAP-based analysis with OFS-matched cutoffs and observe an average 56.7% overlap (range 44.4%–69.2%) between GRAS-selected features and top-ranked SHAP features. Overall, GRAS shows scalable, statistically validated performance and selects business-relevant features, making it a practical choice for churn management pipelines where predictive quality and explainability are required.