Cost-Optimized Customer Churn Prediction System

Overview

This project implements an end-to-end machine learning system to predict customer churn and optimize the prediction threshold based on business cost.

Traditional churn models use a default probability threshold of 0.5 to classify customers. However, in real business scenarios, the cost of misclassification is not equal.

This system identifies the optimal threshold that minimizes business loss by considering:

• False Negative cost (missing a churn customer)
• False Positive cost (wrongly flagging a loyal customer)

The trained model is deployed as a FastAPI REST API and hosted publicly so predictions can be accessed via HTTP requests.

The system includes a trained machine learning model, cost-optimized decision threshold, and a production-style REST API deployed on the cloud.

---

Live Demo

API Endpoint

https://churn-prediction-api-7zca.onrender.com/

Interactive API Docs (Swagger UI)

https://churn-prediction-api-7zca.onrender.com/docs

The Swagger interface allows users to test the model directly from the browser.

You can directly send prediction requests and view responses using the interactive Swagger interface.

---

Problem Statement

Customer churn prediction helps businesses identify customers who are likely to stop using their service.

However, prediction errors have different business consequences:

Prediction Type	Impact
False Negative	Losing a customer without intervention
False Positive	Offering retention incentives unnecessarily

Since losing a customer is more expensive, this project optimizes the prediction threshold using cost-based evaluation instead of a fixed threshold.

Example business cost:

• False Negative Cost = 5000
• False Positive Cost = 500

The model selects the threshold that minimizes total expected cost.

---

System Architecture

Client Request
      │
      ▼
FastAPI REST API
      │
      ▼
ML Model (model.joblib)
      │
      ▼
Threshold Optimization Logic
      │
      ▼
Prediction Response (JSON)

---

Project Structure

cost-optimized-churn-ml-system
│
├── api/
│   └── app.py
│
├── src/
│   ├── preprocessing.py
│   ├── train_model.py
│   └── predict.py
│
├── models/
│   ├── model.joblib
│   └── threshold.json
│
├── metrics/
│
├── data/
│
├── requirements.txt
└── README.md

---

Machine Learning Pipeline

The system follows a structured ML pipeline:

1. Data preprocessing
2. Feature engineering
3. Model training
4. Model evaluation
5. Cost-based threshold optimization
6. Deployment through FastAPI

---

Models Used

The following machine learning models were trained and evaluated:

• Logistic Regression
• Random Forest

Model selection was performed using metrics such as:

• ROC-AUC
• Precision-Recall performance
• Business cost optimization

The final model and threshold are stored as serialized artifacts.

---

Technologies Used

• Python
• Pandas
• Scikit-learn
• FastAPI
• Uvicorn
• Git & GitHub
• Render (Cloud Deployment)

---

API Usage

Health Check Endpoint

To verify that the API service is running:

GET /health

Example:

https://churn-prediction-api-7zca.onrender.com/health

Response:

{
 "status": "ok"
}

---

API Endpoints

Method	Endpoint	Description
GET	/	API information
GET	/health	Service health check
POST	/predict	Predict customer churn

---

Request Example

{
 "gender": "Female",
 "SeniorCitizen": 0,
 "Partner": "Yes",
 "Dependents": "No",
 "tenure": 5,
 "PhoneService": "Yes",
 "MultipleLines": "No",
 "InternetService": "DSL",
 "OnlineSecurity": "No",
 "OnlineBackup": "Yes",
 "DeviceProtection": "No",
 "TechSupport": "No",
 "StreamingTV": "No",
 "StreamingMovies": "No",
 "Contract": "Month-to-month",
 "PaperlessBilling": "Yes",
 "PaymentMethod": "Electronic check",
 "MonthlyCharges": 75.2,
 "TotalCharges": 350.5
}

---

Response Example

{
 "churn_probability": 0.62,
 "threshold_used": 0.147,
 "prediction": 1
}

Where:

• churn_probability → likelihood of customer churn
• threshold_used → optimized decision threshold
• prediction → final churn classification (0 or 1)

---

Quick Test Using Curl

curl -X POST "https://churn-prediction-api-7zca.onrender.com/predict" \
-H "Content-Type: application/json" \
-d '{
"gender":"Female",
"SeniorCitizen":0,
"Partner":"Yes",
"Dependents":"No",
"tenure":5,
"PhoneService":"Yes",
"MultipleLines":"No",
"InternetService":"DSL",
"OnlineSecurity":"No",
"OnlineBackup":"Yes",
"DeviceProtection":"No",
"TechSupport":"No",
"StreamingTV":"No",
"StreamingMovies":"No",
"Contract":"Month-to-month",
"PaperlessBilling":"Yes",
"PaymentMethod":"Electronic check",
"MonthlyCharges":75.2,
"TotalCharges":350.5
}'

---

Key Skills Demonstrated

• Machine Learning Model Development
• Cost-Sensitive Decision Threshold Optimization
• FastAPI REST API Development
• Model Serialization and Deployment
• Cloud Deployment using Render
• End-to-End ML Pipeline Design

---

Future Improvements

Possible extensions for this project include:

• Model monitoring and logging
• CI/CD pipeline for automatic deployment
• Model explainability using SHAP
• Automated retraining pipeline
• Docker containerization

---

Author

Raja Artificial Intelligence & Machine Learning Student Francis Xavier Engineering College

GitHub: https://github.com/Raja-ML-22