Telco-Churn-Prediction-With-Explainability

Overview

This project aims to predict customer churn in the telecommunications industry using advanced ensemble learning techniques and implement personalized retention strategies based on these predictions. The project emphasizes model explainability, providing clear insights into the factors driving customer churn predictions.

Features

• Churn Prediction: Predicts which customers are at risk of leaving using ensemble learning models.
• Cost-Sensitive Learning: Models take into account the financial implications of churn, prioritizing high-cost churns.
• Model Explainability: Utilizes LIME and Counterfactual Explanations to provide transparency into model predictions.
• Personalized Retention Strategies: Implements customized interventions to reduce churn based on predictive insights.
• Interactive Streamlit Application: A user-friendly interface to interact with the model, visualize predictions, and understand the rationale behind them.

Data

The project uses the Telco Customer Churn dataset from Kaggle.

Modeling Approach

Ensemble Learning Models:

• Random Forest
• AdaBoost
• Stacking Classifier

These models are selected for their robustness and ability to handle complex patterns in the data. Hyperparameter tuning was performed using Random Search to optimize model performance.

Cost-Sensitive Learning:

A cost-sensitive approach was applied to minimize the financial impact of churn, ensuring that predictions prioritize high-risk customers who contribute significantly to revenue.

Explainability

LIME (Local Interpretable Model-Agnostic Explanations):

LIME is used to explain individual model predictions, highlighting which features contribute most to the churn prediction for each customer.

Counterfactual Explanations:

Counterfactual explanations show how minimal changes in customer features could alter the churn prediction, providing actionable insights for retention strategies.

Results

• Accuracy: The Stacking Classifier achieved an accuracy of 88.83%.
• AUC Score: The model attained an AUC score of 0.88.
• Churn Reduction: Implemented strategies based on model predictions resulted in a 15% reduction in churn rates.

Installation

1. Clone the repository:

   git clone https://github.com/OvaizAli/Telco-Churn-Prediction-With-Explainability.git
   

2. Navigate to the project directory:

   cd Telco-Churn-Prediction-With-Explainability
   

3. Install the required dependencies:

   pip install -r requirements.txt
   

Streamlit Application

The project includes a Streamlit application that allows users to interact with the model predictions and visualize the explanations.

To run the Streamlit app:

cd streamlit_app
streamlit run app.py

The app will be accessible at https://telco-churn-prediction-with-explainability.streamlit.app/.

Acknowledgments

• Team Members: Haoyu Wang, Anshika Sharma, Priya Mandyal, Kavya Bhojani
• Professor: Ga Wu
• TAs: Sigma, Sher Badsha

Special thanks to everyone involved for their valuable contributions and support throughout this project.