Anticipating maintenance in telecom installation processes

Abstract

Improving customer experience is crucial in any industry, especially in telecommunications, where competition is a constant factor. Today, all telecommunications companies rely on the massive amount of data generated daily to get to know the customer or study their behavior and thus create new effective strategies for their business. Within the most varied user experiences, the process of installing new services can be an event that raises doubts about their operation, degrade the user experience, or, in extreme cases, lead to maintenance interventions. Therefore, the use of advanced predictive models that can predict such occurrences become vital. With this, the company can anticipate the cases that will be problematic and reduce the number of negative experiences. The main objective of this work is to create a predictive model that, through all the available data history, can predict which customers will contact the customer service with problems derived from the installation process and have a following maintenance intervention. After analyzing an unbalanced dataset with approximately 560K entries from a Portuguese telecommunications company, and resorting to the CRISP-DM methodology for modeling, the best results were found with LightGBM which obtained an AUPRC of 0.11 and AUROC of 0.62. The best trade-off between precision and recall was found with a threshold model of 0.43 in order to maximize recall while still avoiding a large number of false negatives.