MACHU-PICCHU: Machine Learning for Predictive and Causal modelling of Churn

Project Overview:

Telecom companies lose more than 30% of customers annually as a result of customer churn in the US and Europe. It has been estimated that acquiring a new customer is approximately eight times more expensive than that of retaining an existing customer. Churn detection and prevention is therefore particularly relevant for a company like Orange Belgium, one of the major telecommunication market actors in Belgium. The Orange Belgium data science team has already got significant experience in the churn detection domain and has set up a churn prevention framework where historical data are used for building predictive models of customers most likely to churn. At the current stage, a limitation of the current approach is the lack of explanation and interpretability of the churn prediction model.

This PhD project focuses on the interpretability of Orange churn predictive models by assuming that an important step forward may derive from the adoption of causal inference techniques. Those techniques aim to identify, within the set of customer variables, the ones which, once manipulated, might lead to a reduction of the churn risk. Since it is essential not only to understand when a customer will churn but also why, we propose the adoption of computational causal analysis strategies which go beyond the scope of conventional descriptive and predictive statistics.



Funding: Applied PhD, INNOVIRIS

Duration: January 2020 – December 2023