Machine Learning Models Based on Geospatial Data for Customer Churn Prediction in ISP

Abstract

The purpose of this study is to address the problem of customer churn in the telecommunications sector, specifically in an Internet Service Provider (ISP) operating in Ecuador, through a machine learning model and the application of geospatial analysis. To carry it out, it was necessary to apply class balancing techniques, since the dataset presented a smaller number of customers in abandonment, in contrast to those who remain in active service. In addition, customer segmentation was carried out using K-Means. In terms of churn prediction, several classification models were evaluated and the one with the best performance was selected. Based on their predictions, geospatial analysis was applied to examine the territorial distribution of customers, identify patterns and enable the development of more effective retention strategies. To evaluate each model, ROC-AUC and recall metrics were applied, where Random Forest in combination with Random Downsampling presented the best performance. While, through segmentation, four groups of customers with similar characteristics were identified. In conclusion, this study demonstrates that integrating machine learning and geospatial analysis is an effective combination for predicting customer churn in the telecommunications sector. The combination of Random Forest with data balancing techniques, together with customer segmentation using K-Means, resulted in a robust and accurate model.