💳 Segmentation of Credit Card Customers

An unsupervised machine learning project that segments ~9,000 active credit card holders into distinct behavioural clusters — enabling banks to design targeted marketing strategies and personalised financial products.

---

📌 Problem Statement

A bank's marketing team needs to understand who their credit card customers really are — not just as account holders, but as distinct behavioural groups with different spending habits, payment patterns, and financial needs.

With 9,000 active customers and 18 behavioural variables, this project applies unsupervised ML techniques to:

1. Engineer intelligent KPIs from raw transactional data
2. Reduce dimensionality using Factor Analysis to remove multicollinearity
3. Cluster customers into meaningful behavioural segments
4. Profile each segment with detailed characteristics
5. Recommend targeted marketing strategies for each cluster

---

🎯 Business Impact

Stakeholder	Value Delivered
Marketing Team	Run personalised campaigns tailored to each customer segment
Product Team	Design credit products (cashback, travel, premium) for specific segments
Risk Team	Identify high-risk segments (heavy cash advance users, minimum payment payers)
Operations	Prioritise service improvements for high-value but low-satisfaction segments
Executive Leadership	Data-driven strategic insights on customer base composition

According to Bain & Company, companies with excellent customer segmentation strategies see up to 10% profit increase over a 5-year period compared to those without.

---

📊 Dataset

• Source: Kaggle — Credit Card Dataset for Clustering
• File: CC_GENERAL.csv
• Size: ~9,000 active credit card holders
• Period: Last 6 months of behavioural data
• Granularity: Customer level with 18 behavioural variables

Data Dictionary

Variable	Description
CUST_ID	Unique credit card holder ID
BALANCE	Monthly average balance (based on daily averages)
BALANCE_FREQUENCY	Ratio of last 12 months with a balance (0–1)
PURCHASES	Total purchase amount in last 12 months
ONEOFF_PURCHASES	Total one-off (single) purchase amount
INSTALLMENTS_PURCHASES	Total instalment purchase amount
CASH_ADVANCE	Total cash advance amount taken
PURCHASES_FREQUENCY	% of months with at least one purchase
ONEOFF_PURCHASES_FREQUENCY	Frequency of one-off purchases
PURCHASES_INSTALLMENTS_FREQUENCY	Frequency of instalment purchases
CASH_ADVANCE_FREQUENCY	Frequency of cash advances
CASH_ADVANCE_TRX	Average amount per cash advance transaction
PURCHASES_TRX	Average amount per purchase transaction
CREDIT_LIMIT	Customer's credit limit
PAYMENTS	Total payments made to reduce statement balance
MINIMUM_PAYMENTS	Total minimum payments made
PRC_FULL_PAYMENT	Percentage of months with full payment
TENURE	Number of months as a credit card customer

---

🏗️ Project Architecture

Segmentation-of-Credit-Card-Customers/
├── CC_GENERAL.csv               # Raw customer behavioural dataset
├── Credit_new_cust.xlsx         # New customer profiles for segment assignment
├── Profiling_output.csv         # Cluster profiling results
├── python_FA.xls                # Factor Analysis output
├── pandas_profiling.html        # Auto-generated EDA report (open in browser)
├── Segmentation of Credit Card Customers.ipynb   # Full pipeline notebook
├── LICENSE
└── README.md

---

🔍 Full Analytical Pipeline

Step 1 — Advanced Data Preparation & KPI Engineering

Raw transactional variables alone don't tell the full story. Intelligent KPIs are engineered to capture customer behaviour ratios that reveal true financial habits:

Engineered KPI	Formula	Business Meaning
monthly_avg_purchase	PURCHASES / TENURE	Average monthly spending rate
monthly_avg_cash_advance	CASH_ADVANCE / TENURE	Monthly reliance on cash advances
limit_usage_ratio	BALANCE / CREDIT_LIMIT	How much of credit limit is used
payments_to_min_ratio	PAYMENTS / MINIMUM_PAYMENTS	Payment discipline (higher = better)
purchase_type_ratio	ONEOFF / INSTALLMENTS	Spending style — impulsive vs planned
cash_advance_ratio	CASH_ADVANCE / PURCHASES	Reliance on cash vs purchases

Step 2 — Exploratory Data Analysis (EDA)

• Generated full statistical profiling report using pandas-profiling (pandas_profiling.html)
• Analysed distributions of all 18 variables — most spending variables are highly right-skewed
• Identified significant missing values in MINIMUM_PAYMENTS and CREDIT_LIMIT — imputed using median
• Correlation analysis revealed multicollinearity between purchase frequency variables — motivating Factor Analysis

Step 3 — Dimensionality Reduction: Factor Analysis

With 18 correlated variables, direct clustering leads to the Curse of Dimensionality. Factor Analysis groups correlated variables into latent factors that represent underlying behavioural dimensions:

# Factor Analysis for variable reduction
from factor_analyzer import FactorAnalyzer
fa = FactorAnalyzer(n_factors=optimal_factors, rotation='varimax')
fa.fit(scaled_data)
factor_scores = fa.transform(scaled_data)

Output saved to python_FA.xls for detailed factor loading analysis.

Identified Latent Factors (example):

• Factor 1 — Purchase Activity: Groups purchase frequency, purchase amount, transaction count
• Factor 2 — Cash Advance Behaviour: Groups cash advance amount, frequency, transactions
• Factor 3 — Payment Discipline: Groups full payment %, payments to minimum ratio

Step 4 — Clustering: K-Means Segmentation

from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler

# Scale factor scores
scaler = StandardScaler()
scaled_factors = scaler.fit_transform(factor_scores)

# Optimal k using Elbow Method
inertias = []
for k in range(2, 11):
    km = KMeans(n_clusters=k, random_state=42)
    km.fit(scaled_factors)
    inertias.append(km.inertia_)

Optimal number of clusters: 4 (determined via Elbow Method + Silhouette Score)

Step 5 — Cluster Profiling & Strategic Insights

Each cluster is profiled on key variables to understand who they are and what they need:

Cluster	Profile Name	Characteristics	Recommended Strategy
0	💤 Inactive / Low Engagers	Low balance, low purchases, low cash advance	Re-engagement campaigns, introductory cashback offers
1	💰 High Spenders	High purchases, high credit limit, frequent transactions	Premium rewards, travel cards, loyalty programs
2	🏧 Cash Advance Reliant	High cash advance, low purchases, high interest burden	Financial wellness programs, personal loan conversion
3	📦 Instalment Shoppers	High instalment purchases, consistent payment behaviour	EMI offers, buy-now-pay-later products, retail partnerships

Note: Exact cluster labels depend on the final model output in the notebook.

---

📈 Key Findings

• ~30% of customers rely heavily on cash advances — a high-risk, high-interest segment that the bank should either convert to personal loans or offer financial planning support
• High spenders (Cluster 1) have the highest credit limits but also the highest balance-to-limit ratios — prime candidates for premium card upgrades
• Full payment behaviour (PRC_FULL_PAYMENT) is the single strongest differentiator between financially healthy and at-risk customer segments
• Instalment shoppers show the most consistent, predictable behaviour — lowest churn risk and best candidates for cross-sell

---

🛠️ Tech Stack

Category	Tools
Language	Python 3.8+
Data Processing	Pandas, NumPy
EDA	pandas-profiling
Visualisation	Matplotlib, Seaborn
Dimensionality Reduction	Factor Analyzer, PCA (Scikit-learn)
Clustering	Scikit-learn (K-Means)
Cluster Validation	Elbow Method, Silhouette Score
Output	Excel (openpyxl) for profiling reports
Notebook	Jupyter Notebook

---

🚀 Running Locally

# Clone the repository
git clone https://github.com/vicky60629/Segmentation-of-Credit-Card-Customers.git
cd Segmentation-of-Credit-Card-Customers

# Install dependencies
pip install -r requirements.txt

# Launch the notebook
jupyter notebook "Segmentation of Credit Card Customers.ipynb"

# View the EDA report
open pandas_profiling.html   # or double-click to open in browser

---

💡 Key Learnings & Future Improvements

What worked well:

• Factor Analysis before clustering significantly reduced noise and multicollinearity — clusters are more interpretable than raw K-Means on 18 variables
• Engineering KPI ratios (limit usage, payment discipline) revealed behavioural patterns invisible in raw variables
• Profiling output (Profiling_output.csv) makes cluster insights shareable with non-technical stakeholders

Future enhancements:

• Try DBSCAN or Gaussian Mixture Models for non-spherical cluster shapes
• Add Hierarchical Clustering + Dendrogram for alternative segmentation view
• Build interactive Tableau / Plotly Dash dashboard for cluster exploration
• Apply RFM Analysis (Recency, Frequency, Monetary) alongside behavioural clustering
• Extend to real-time segment assignment for new customers using Credit_new_cust.xlsx pipeline
• Add SHAP / interpretability layer to explain individual customer segment assignment
• Track experiments with MLflow

---

🧠 Why Unsupervised Learning Here?

Unlike supervised learning, there are no predefined labels for customer segments — we don't know upfront how many types of customers exist or what they look like. Unsupervised learning lets the data speak for itself, revealing natural groupings that a business analyst might never discover manually from 9,000 rows and 18 columns.

This is exactly the kind of analysis that drives real decisions at banks: product design, marketing spend allocation, and risk management — all informed by who customers actually are rather than assumptions.

---

👨‍💻 About the Author

Vicky Gupta — Data Engineering Analyst @ Accenture (4.5 years) | Aspiring Data Scientist

Passionate about unsupervised ML, customer analytics, and building data products that drive real business decisions. Experienced in PySpark, ETL pipelines, and end-to-end ML workflows.

🔗 LinkedIn | GitHub

📧 vg60629@gmail.com

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

⭐ Found this useful? Star the repository — it helps others in the community discover it!