Situation

A client in the telecommunications sector was experiencing significant client churn, leading to a substantial loss in market share.

Task

The client required a reliable predictive system to identify customers at risk of churning, enabling them to offer targeted special deals and retain those customers. Moreover, the client requested a system with a large degree of interpretability to accurately identify causes of churning.

Action

I conducted an exploratory data analysis to identify anomalies in the dataset. Subsequently, I developed a comprehensive system to classify customers into "low risk" and "high risk" categories. This system included:

• Customer data import via spreadsheet format or SQL database.

• Data cleaning to ensure accuracy and consistency.

• Training ensembles of decision trees.

• Cross-validation to ensure model robustness.

• Performance evaluation and ongoing  monitoring using MLflow.

• Model serving via FastAPI to facilitate integration with the client’s existing infrastructure.

• Appealing visualization of churn likelihoods of customers, most important input features as well as underlying ML models via gradio

• Identification of causal relationships between input data and churn likelihood

• Deployment in the client's Kubernetes environment for scalability.

Result

The client could automatically identify and approach at-risk customers through a targeted email campaign. This proactive approach significantly reduced customer churn and helped the client maintain and eventually secure market shares.