Predicting Effectiveness of Churn Prevention Measures
Losing customers, also referred to as churning, is something that any company wants to prevent. While it is interesting to know how likely a specific customer will churn, it is more useful to know which countermeasure, such as a discount or an additional service, would prevent a user from churning. Also, instead of knowing which countermeasure works well in general, we want to determine the best countermeasure for each individual customer.
Since we are interested in the causal relationship between countermeasure and churn, we need to go beyond analyzing correlations. For example, did the customer churn because he got the wrong countermeasure, or did he get the countermeasure because he was likely to churn anyway?
Furthermore, the feedback from previous countermeasures is limited to the set of actually executed countermeasures. This phenomenon, referred to as bandit feedback, does not provide information about the alternative, potentially more effective countermeasure. As the model selects the applied countermeasure, our data only contains feedback about the predicted countermeasures. This data generation process causes very strong selection biases, and introduces an extra challenge of balancing between exploiting predicted associations and exploring unknown associations.
In this talk, I will demonstrate how to obtain unbiased predictions of the effect of countermeasures, which can be used to select the optimal countermeasures, while balancing between exploiting and exploring. This will be done by combining deep learning and bayesian modelling in a custom setup.
Gerben Oostra is a machine learning engineer that loves to solve complex problems. Originally graduated on algorithmic approaches in operations research, he switched to data science challenges about 10 years ago. Since then he has been applying data science and data engineering in various companies, amongst which ING and Vodafone/Ziggo. He therefore gained data science experience on a variety of challenges, including social network analysis, time series forecasting, anomaly (fraud) detection, natural language processing and bayesian modelling. More recently he has focused on solving churn cases, allowing him to combine his enthusiasm about both Bayesian modelling and time-series on a real life problems.