RFM Customer Segmentation - ML-Driven Approach

End-to-end customer analytics pipeline that segments 4,290 customers from £6.7M in retail transactions using rule-based RFM scoring and unsupervised clustering. Includes a systematic comparison of K-Means vs Gaussian Mixture Model across multiple preprocessing strategies.

---

Business Impact

Metric	Value
Total revenue analyzed	£6.7M across 394k transactions
Champions identified	1,127 customers generating £4.4M (65%)
Revenue at risk	£575k from 512 at-risk customers
Actionable segments	10 with tailored marketing strategies

---

What Makes This Project Different

Most RFM analyses on this dataset stop at K-Means with default settings. This project goes further:

1. Algorithm comparison — K-Means vs GMM, systematically benchmarked
2. Preprocessing comparison — log-transform vs Yeo-Johnson power transform
3. Statistical validation — Hopkins statistic (0.956) proves data is clusterable before running algorithms
4. Production-ready code — Reusable RFMPipeline class, not just a notebook
5. Tested and automated — 61 unit tests, GitHub Actions CI across Python 3.10-3.12
6. Iterative development — v1.0 baseline, then v2.0 with multi-algorithm comparison via documented PR

---

Quick Start

git clone https://github.com/leelesemann-sys/rfm-customer-segmentation.git
cd rfm-customer-segmentation
pip install -r requirements.txt

python run_pipeline.py                    # Run full pipeline with defaults
python run_pipeline.py --k 5             # Try different cluster counts

---

Results

RFM Segments (Rule-Based)

Priority	Segment	Customers	Revenue	Recommended Action
High	Champions	1,127	£4.4M	VIP programs, loyalty rewards
High	At Risk	453	£508k	Win-back campaigns, 20% discount
High	Can't Lose Them	59	£67k	Personal outreach, account managers
Medium	Loyal Customers	802	£994k	Upselling, cross-sell
Medium	New Customers	136	£44k	Onboarding, next purchase incentive
Low	Lost	798	£294k	Low-cost reactivation only
Low	Hibernating	399	£179k	Mass email campaigns

Algorithm Comparison

Algorithm	Transform	Silhouette	Davies-Bouldin	Winner?
K-Means	log	0.380	0.857	Best
K-Means	Yeo-Johnson	0.338	1.019
GMM	Yeo-Johnson	0.197	1.768
GMM	log	0.112	1.851

Key insight: Contrary to Shobayo et al. (2023) who found GMM superior (Silhouette 0.80 vs 0.62), K-Means outperforms GMM on this dataset. The log-transform makes RFM features approximately spherical, which is exactly what K-Means assumes. GMM's flexibility (elliptical clusters) adds complexity without improving separation.

K-Means Clusters (K=4)

Cluster	Size	Avg. Recency	Avg. Purchases	Avg. Spend
Inactive	921	260 days	1	£356
Regular	1,341	59 days	1	£359
VIP Regulars	1,434	47 days	4	£1,442
Super VIPs	594	19 days	15	£6,457

Dashboards

---

Methodology

Raw Data (541k rows)
    │
    ├── 1. Data Cleaning ──────────── 394k transactions retained (72.7%)
    ├── 2. RFM Aggregation ────────── 4,290 customer profiles
    ├── 3. Quintile Scoring ───────── R/F/M scores (1-5 scale)
    ├── 4. Rule-Based Segments ────── 10 business segments
    ├── 5. Hopkins Statistic ──────── 0.956 (clustering validated)
    ├── 6. Preprocessing ──────────── log-transform vs Yeo-Johnson
    ├── 7. Algorithm Comparison ───── K-Means vs GMM (4 combinations)
    └── 8. Best Model ─────────────── K-Means + log (Silhouette: 0.380)

---

Tech Stack

Category	Tools
Language	Python 3.11
Data	pandas, numpy
ML	scikit-learn (K-Means, GMM, Hopkins, Yeo-Johnson)
Visualization	matplotlib, seaborn
Testing	pytest (61 tests), pytest-cov
CI/CD	GitHub Actions (Python 3.10, 3.11, 3.12)

---

Project Structure

rfm-customer-segmentation/
├── src/
│   ├── __init__.py
│   └── rfm_pipeline.py               # Reusable pipeline class (K-Means, GMM, Hopkins)
├── notebooks/
│   ├── 01_data_exploration.ipynb      # EDA & data cleaning
│   └── 02_rfm_analysis.ipynb         # RFM scoring & clustering
├── tests/
│   ├── conftest.py                    # Shared test fixtures (50 synthetic customers)
│   └── test_pipeline.py              # 61 unit tests across 10 test classes
├── visualizations/                    # 7 publication-ready PNGs
├── data/
│   └── online_retail_clean.csv.zip   # Cleaned dataset (394k transactions)
├── run_pipeline.py                    # CLI entrypoint (full pipeline + all visualizations)
├── .github/workflows/test.yml         # CI: tests + coverage badge
└── requirements.txt

---

Dataset

Source: UCI Machine Learning Repository — Online Retail Period: Dec 2010 -- Dec 2011 (12.4 months) Size: 541,909 transactions | 4,290 unique customers | UK-based (89.1%)

---

Version History

Version	What changed	PR
v1.0	Baseline: RFM + K-Means, 36 tests, CI	--
v2.0	+GMM, +Yeo-Johnson, +Hopkins, 61 tests	#1

---

Future Enhancements

• Predictive CLV model (Random Forest / XGBoost)
• Churn prediction classifier
• Real-time segmentation API (FastAPI)
• Interactive dashboard (Streamlit or Power BI)

---

Author

Lee Christian Lesemann Azure AI Engineer | Customer Analytics Consultant Previous: Sanofi, CSL Behring, Abbott, Teva Pharmaceuticals, IQVIA

---

License

MIT License -- see LICENSE for details