💳 Customer Segmentation Platform – End-to-End ML System

Production-ready customer segmentation system using unsupervised machine learning to identify behavioral clusters in credit card data.

📌 Business Context

Banks and fintech companies need to:

• Identify high-value customers
• Detect risky spending behavior
• Personalize credit offers
• Optimize marketing spend
• Improve customer retention

This project builds a scalable ML segmentation pipeline to solve that.

🧠 System Architecture

🔎 1️⃣ Data Engineering

Preprocessing

• Missing value imputation
• Outlier detection
• Log transformation for skewed features
• Standard scaling

Feature Engineering

• Credit utilization ratio
• Purchase frequency metrics
• Cash advance ratio
• Payment consistency index

🤖 2️⃣ Model Development

Algorithms Evaluated

Model	Silhouette Score	Stability
K-Means	0.42	High
Hierarchical	0.39	Medium
DBSCAN	0.21	Low

Final model: K-Means (k=4)

Cluster Selection Strategy

• Elbow Method
• Silhouette Analysis
• Business interpretability validation

🏷 3️⃣ Identified Customer Segments

🥇 Premium High Spenders

High balance, high purchases, strong credit limits.

🥈 Stable Customers

Moderate spending, consistent payments.

🥉 Cash Advance Heavy Users

High withdrawals, potential financial risk.

💤 Low Engagement Users

Low activity, low revenue contribution.

📊 4️⃣ Evaluation Strategy

• Silhouette Score
• Inertia
• PCA visualization
• Cluster stability tests
• Feature importance per cluster

🏗 5️⃣ End-to-End ML Pipeline

Implemented using sklearn Pipeline:

Pipeline([
    ('scaler', StandardScaler()),
    ('pca', PCA(n_components=0.90)),
    ('cluster', KMeans(n_clusters=4))
])

Pipeline supports:

• Re-training
• Batch scoring
• Real-time inference
• Reproducibility

🌐 6️⃣ Deployment Architecture

🔹 REST API (FastAPI)

Endpoints:

• /predict → Assign cluster
• /health → Check service
• /retrain → Trigger retraining

🔹 Dockerized Service

docker build -t segmentation-api .
docker run -p 8000:8000 segmentation-api

📊 7️⃣ Monitoring & MLOps

• Cluster drift detection
• Distribution monitoring
• Re-clustering trigger logic
• Versioned models
• Experiment tracking (optional MLflow)