A hybrid approach for predicting and preventing churn

As part of my MBA in Data Science & Analytics, I proposed a hybrid approach using unsupervised and supervised models for predicting customer churn from a telecom company.

Model training:

The unsupervised pipeline includes dimensionality reduction using Factor analysis of mixed data (FAMD) followed by K-means clustering, which allowed us to develop better strategies based on the cluster patterns.

The supervised model pipeline includes Yeo-Johnson transformation to the numerical features, one-hot encoding to the categorical features, and weighted logistic regression for churn classification.

All code was developed using best MLOps practices, open-source pipelines, and proper validation, therefore avoiding data leakage.

Model deployment:

The model was deployed using a web app for interactive "what if" or "batch" predictions and a REST API with proper documentation.

Links:

• Web app
  • Web app deployed on the Streamlit cloud
  • Docker image repository
  • Code repository
• REST API
  • REST API deployed on GCP App Engine (framework: FastAPI)
  • Docker image repository
  • Code repository