🛍️ Customer Segmentation for E-Commerce

Unsupervised Machine Learning project that segments retail customers into actionable personas using RFM analysis and clustering (K-Means, Hierarchical, DBSCAN).

---

📌 Project Overview & Problem Statement

An e-commerce company is running the same marketing strategy for every customer — same emails, same discounts, same ads — regardless of whether the customer is a high-spending loyalist or someone who hasn't ordered in months. This is expensive and ineffective.

Goal: Use unsupervised machine learning to segment the customer base into a small number of statistically distinct, business-interpretable groups based on income, spending behavior, purchase frequency, and recency — so marketing can design a tailored strategy per segment instead of one blurry strategy for everyone.

This project covers the full data science lifecycle: data sourcing, cleaning, EDA, feature engineering (RFM), clustering, model comparison, evaluation, business storytelling, and an interactive dashboard.

---

📊 Dataset

This project uses a hybrid real + engineered dataset, which is disclosed transparently below.

Real data (from Kaggle): Mall Customer Segmentation Data by vjchoudhary7 — one of the most widely used public customer-segmentation datasets. It provides 200 real customer records with CustomerID, Gender, Age, Annual Income, and Spending Score. A local copy is stored at data/raw/Mall\_Customers.csv.

Why augment it: The real Kaggle file has no transaction history (no order dates, no region, no category data), so it cannot support RFM analysis or a richer segmentation on its own, and at 200 rows it's small for a robust clustering demo. src/build\_dataset.py expands the 200 real rows to 1,200 (resampling + small realistic jitter that preserves the original distribution) and adds behavioral columns that are simulated conditionally on each customer's real Spending Score / Income — not independent random noise — so the added behavior stays statistically consistent with real signal. The full logic is documented in that file.

Column	Source	Description
CustomerID	Real (Kaggle)	Unique customer identifier
Age	Real (Kaggle, jittered)	Customer age
Gender	Real (Kaggle)	Male / Female
AnnualIncome\_INR\_K	Real (Kaggle, jittered)	Annual income, thousands
SpendingScore	Real (Kaggle, jittered)	Store-assigned spending score (1-100)
City	Engineered	Simulated region across 10 major Indian cities
PurchaseFrequency	Engineered	Orders/year, conditional on Spending Score
Recency\_Days	Engineered	Days since last order, conditional on Spending Score
Tenure\_Months	Engineered	Months as a customer, conditional on Spending Score
ProductCategoryPreference	Engineered	Preferred product category
AvgOrderValue\_INR	Engineered	Derived from income + spending score
MonetaryValue\_INR	Engineered	Total spend, used as the "M" in RFM
SatisfactionScore	Engineered	1-5 satisfaction rating, conditional on Spending Score

Rows: 1,200 · Columns: 13 · Includes realistic missing values (~2%) and outliers, same as you'd find in a real CRM export.

Want to regenerate it? Run python src/build\_dataset.py.

---

🛠️ Tools & Tech Stack

Category	Tools
Language	Python 3.10+
Data handling	pandas, NumPy
Visualization	matplotlib, seaborn, Plotly
Machine Learning	scikit-learn (K-Means, Agglomerative Clustering, DBSCAN, PCA)
Dashboard	Streamlit
Notebook	Jupyter
Reporting	Markdown / PDF executive summary

---

🧭 Methodology

1. Data Sourcing — Load real Kaggle base data, expand + enrich into the working dataset (src/build\_dataset.py).
2. Data Cleaning — Handle missing values (median/mode imputation), cap outliers (IQR method).
3. EDA — Univariate & bivariate visualizations, correlation heatmap, pairplots, insight write-ups.
4. Feature Engineering — RFM (Recency, Frequency, Monetary) scoring, categorical encoding, StandardScaler feature scaling.
5. Optimal k Selection — Elbow method (inertia) + Silhouette score across k=2 to 10.
6. Clustering — K-Means, Hierarchical (Ward linkage), and DBSCAN, compared head-to-head.
7. Dimensionality Reduction — PCA to 2D/3D for visualization.
8. Cluster Profiling — Named, business-interpretable personas with average-behavior tables.
9. Model Evaluation — Silhouette Score, Davies-Bouldin Index, cluster stability check across resamples.
10. Dashboard — Interactive Streamlit app with filters, KPIs, and segment visualizations.

---

Screenshots

---

💡 Key Insights & Business Impact

• 4 clean, stable customer segments were found: Premium Spenders (21%), Steady Mature Regulars (24%), New Bargain Hunters (28%), and At-Risk Customers (27%).
• K-Means (k=4) was selected as the production model — it matched Hierarchical clustering's structure closely (validating the segments are real) and produced clean, balanced, business-usable groups, unlike DBSCAN, which fragmented the data into one dominant cluster plus many tiny outlier groups despite a higher raw silhouette score.
• At-Risk Customers have relatively high income but the lowest spending score and highest recency (~120 days since last order) — a high-value win-back opportunity, since low income isn't the barrier.
• Cluster stability check (10 reruns on 80% resamples) showed low variance in silhouette score, confirming the segments are robust and not a one-off artifact.
• Business impact: replacing one blanket marketing strategy with four targeted ones (VIP tier, cross-sell nudges, price-sensitive promos, and a win-back campaign) directly targets retention and revenue growth on the highest-opportunity segments.

Full recommendations per segment are in the notebook (Section 15) and reports/executive\_summary.md / .pdf.

---

🚀 How to Run This Project

1. Clone & install dependencies

git clone <your-repo-url>
cd customer-segmentation
pip install -r requirements.txt

2. (Re)generate the dataset

python src/build\_dataset.py

3. Run the analysis notebook

jupyter notebook notebooks/Customer\_Segmentation\_Analysis.ipynb

4. Regenerate all chart images (optional)

python src/generate\_images.py

5. Launch the interactive dashboard

streamlit run app.py

---

📁 Folder Structure

customer-segmentation/
│
├── data/
│   ├── raw/
│   │   └── Mall\_Customers.csv          # Real Kaggle source data (200 rows)
│   ├── customer\_data.csv               # Final working dataset (1,200 rows, real + engineered)
│   ├── customer\_data\_with\_segments.csv # Output: customers labeled with their cluster/segment
│   ├── rfm\_table.csv                   # RFM scores per customer
│   ├── cluster\_profile.csv             # Avg behavior per cluster
│   └── algorithm\_comparison.csv        # K-Means vs Hierarchical vs DBSCAN metrics
│
├── notebooks/
│   └── Customer\_Segmentation\_Analysis.ipynb   # Full, commented, story-driven analysis
│
├── src/
│   ├── build\_dataset.py     # Loads real Kaggle data + builds the enriched working dataset
│   ├── preprocessing.py     # Cleaning, RFM, encoding, scaling utilities
│   ├── clustering.py        # K-Means / Hierarchical / DBSCAN / PCA / evaluation utilities
│   └── generate\_images.py   # Reproduces every chart used in this README/report
│
├── reports/
│   ├── executive\_summary.md   # Business-facing summary (source)
│   └── executive\_summary.pdf  # Business-facing summary (client-ready PDF)
│
├── images/                  # All generated charts (EDA, clustering, profiling)
│
├── app.py                   # Streamlit interactive dashboard
├── requirements.txt
└── README.md

---

🔭 Future Scope

• Swap the engineered behavioral columns for real transaction-level data (order history, timestamps) if/when available, to remove the simulation layer entirely.
• Add a supervised "next best action" model on top of segments (e.g. predict churn probability within the At-Risk segment).
• Track segment migration over time (cohort analysis) to measure whether marketing interventions actually move customers into higher-value segments.
• Deploy the Streamlit dashboard (Streamlit Community Cloud / Docker + cloud hosting) for live stakeholder access.
• Experiment with Gaussian Mixture Models and HDBSCAN as additional clustering baselines.
• Automate the pipeline (Airflow/Prefect) to re-segment customers on a recurring schedule as new data arrives.

---

👤 Author / Contact

Harshit Fartiyal
📧 Email: fartiyalharshit41@gmail.com
🔗 LinkedIn: LinkedIn Profile
💻 GitHub: Harshit06

This project was built as part of an internship/portfolio submission to demonstrate an end-to-end unsupervised machine learning workflow — from raw data to a business-ready dashboard.

---

📄 License

This project is released under the MIT License — free to use, modify, and learn from.