Telco Customer Churn Predictor

This repository contains an end-to-end deep learning pipeline designed to predict customer churn using PyTorch. By analyzing customer demographics, account information, and service usage, the model identifies high-risk customers, allowing businesses to act precisely with retention strategies through a live interactive dashboard.

🚀 Project Overview

• Objective: Predict binary churn (Yes/No) with high precision.
• Model: Multilayer Perceptron (MLP) with Batch Normalization and Dropout.
• Performance: Achieved ~79.74% accuracy on the validation set.
• Interactive Interface: Live Streamlit web application for real-time risk scoring.

🛠️ Step-by-Step Process

1. Step 1: We defined churn as the target variable (Binary Classification) to help the sales team prioritize retention efforts.

2. Step 2 (Data): Cleaned the Kaggle Telco dataset, handling "empty string" errors in TotalCharges.

• Encoded categorical features using LabelEncoder.

• Applied StandardScaler to ensure numerical features (Tenure, Charges) had a mean of 0 and variance of 1.

3. Step 3 (Architecture): Designed a 3-layer MLP architecture (64 -> 32 -> 1).

• Implemented BatchNorm1d for training stability and Dropout to prevent overfitting.

4. Step 4 (Training):

• Used BCEWithLogitsLoss for numerical stability.

• Trained for 30 epochs, reaching a final training loss of 0.4060.

5. Step 5: Monitored accuracy across epochs, starting at 78.54% and peaking at 79.74%.

6. Step 6: Developed an inference.py script for individual customer scoring.

• Deployed a Streamlit dashboard for non-technical stakeholders to perform "What-If" analysis.

📊 Business Intelligence Visualizations

The following visualizations provide context for why the model makes specific predictions:

1. Churn Rate by Contract Type and Gender

• Observation: Customers on Month-to-Month contracts exhibit drastically higher churn rates compared to One-year or Two-year contracts.

• Insight: Contract flexibility is the primary driver of attrition; gender plays a minimal role.

2. Customer Churn Rate by Tenure Group

• Observation: Churn is highest in the 0-1 Year group and decreases significantly as tenure increases.

• Insight: The first 12 months are the "critical period" requiring focused onboarding.

3. Churn Rate by Internet Service and Online Security

• Observation: Fiber Optic users without Online Security are high-risk.

• Insight: Bundling security services with high-speed internet is a strategic retention necessity.

4. Most Commonly Used Payment Methods

• Observation: Electronic Check is the most frequent payment method but is historically linked to higher churn.

• Insight: Incentivizing a move toward Automatic Credit Card or Bank Transfer could reduce involuntary churn.

📁 Repository Structure

• data/: Contains the WA_Fn-UseC_-Telco-Customer-Churn.csv.
• dataset.py: Preprocessing and PyTorch Dataset class.
• model.py: Neural Network architecture.
• train.py: Logic for training and validation loops.
• main.py: Entry point for the full training pipeline.
• inference.py: Script to generate a risk score (0-100) for a single customer.
• app.py: Streamlit web interface for interactive predictions.
• requirements.txt: Environment dependencies.

💻 How to Run

1. Clone the repo and navigate to the directory.

git clone https://github.com/devopspower/telco-customer-churn-predictor.git

3. Install dependencies:

pip install -r requirements.txt

3. Train the model:

python main.py

4. Launch the Interactive Dashboard:

streamlit run app.py