Unsupervised-Learning

A growing collection of hands-on, research-driven projects exploring clustering, segmentation, and pattern discovery.

This repository documents my journey into Unsupervised Machine Learning — a domain focused on discovering hidden structure in unlabeled data. The goal is to learn by doing: experimenting with multiple clustering algorithms, evaluating their performance, comparing behaviors, and understanding where each method succeeds or fails.

The repo will expand over time with multiple practice projects, each inside its own sub-directory. Currently, it includes work on:

• Customer segmentation using multiple clustering methods
• Online retail segmentation using RFM + KMeans

Each project is structured in a way similar to an applied ML research notebook — clear EDA → thoughtful preprocessing → algorithm exploration → evaluation → interpretation.

---

📁 Repository Structure

Unsupervised-Learning/
│
├── Online retail cluster analysis with Kmeans.ipynb
│     └── Full RFM-based customer segmentation using KMeans  
│
├── Customer Cluster Analysis.ipynb
│     └── Multi-algorithm clustering exploration (KMeans, DBSCAN, Agglomerative, Divisive)
│
└── (future folders)
       ├── market-segmentation/
       ├── anomaly-detection/
       ├── dimensionality-reduction/
       └── kaggle-practice-projects/

---

1. Online Retail Customer Segmentation (RFM + KMeans)

Notebook: Online retail cluster analysis with Kmeans.ipynb

Dataset

• Kaggle “Online Retail II”

• 525k transactions

• Includes: invoice, product, quantities, prices, customer IDs, timestamp, country

Data Cleaning & Important Decisions

• Removed missing Customer ID records (as segmentation requires identifiable users)
• Identified and removed returns by excluding invoices starting with “C”
• Cleaned all negative quantities and prices
• Engineered total monetary spend per row: Total = Quantity × Price

These cleaning steps prevent misleading RFM metrics.

RFM Feature Engineering

Calculated per customer:

Feature	Meaning
Recency	Days since last purchase
Frequency	Number of unique invoices
Monetary	Total spending

Applied:

• Log transformation (to reduce skew)
• Standardization (for KMeans distance stability)

KMeans Clustering

• Used k-means++ initialization
• Tested k = 2–10 using the Elbow Method
• Optimal k ≈ 4

Evaluation

• Silhouette Score: ~0.33
• Davies–Bouldin: ~1.009 (Healthy but imperfect — typical for retail behavior data.)

Cluster Insights

Cluster	Size	Customer Behavior
0	779	VIP segment — high monetary, frequent shoppers
1	1403	High-recency (inactive), low frequency, low spend
2	1185	Mid-spend, moderate recency, regular customers
3	947	Low frequency, moderate recency

Visualizations

• Cluster heatmap
• Recency–Monetary scatter
• Cluster size distribution
• Pie charts of proportions

Model Saved

kmeans_model.joblib

---

2. Multi-Algorithm Cluster Analysis (Customer Attributes Dataset)

Notebook: Customer Cluster Analysis.ipynb

Dataset

Kaggle customer segmentation data with:

• Sex
• Marital status
• Age
• Education
• Income
• Occupation
• Settlement size

2000 rows, all numeric categories.

EDA & Distributions

Explored:

• Skewness and kurtosis of Age and Income
• Category count distributions
• Correlation between numeric & ordinal features
• Pairplots for deeper intuition

Preprocessing

• Removed ID column
• Separated numeric vs categorical features
• No missing values
• Standardization performed inside each clustering pipeline

---

Clustering Algorithms Implemented

1. KMeans

• Pipeline: StandardScaler → KMeans

• Baseline k=3

• Silhouette ≈ 0.017

  • (Low: dataset has overlapping clusters)

2. DBSCAN

• Tuned using k-distance graph

• Silhouette ≈ –0.687

  • High density variance → DBSCAN struggled

3. Agglomerative Clustering

• Linkage: Ward
• Performs best among tested algorithms
• Silhouette ≈ 0.113

4. Divisive Clustering (Custom Implementation)

• Wrote a custom hierarchical divisive clustering class
• Repeatedly splits largest cluster using KMeans
• Silhouette ≈ 0.024

---

Model Comparisons

Silhouette Score (↑ better):

KMeans = 0.017 DBSCAN = –0.687 Agglomerative = 0.113 Divisive = 0.024

Davies–Bouldin Index (↓ better):

Agglomerative = 1.924 Divisive = 1.805 KMeans = 2.056 DBSCAN = 202.9

Best overall: Agglomerative Clustering

---

Goals of the Repository

Build practical intuition

Learn how different clustering algorithms behave on real-world messy datasets.

Research-oriented experimentation

Evaluate models using Silhouette, Davies–Bouldin, inertia, and interpretability.

Develop a reusable structure

Future practice notebooks & Kaggle projects will follow the same pattern:

1. EDA
2. Preprocessing
3. Feature engineering
4. Multiple clustering methods
5. Evaluation & comparison
6. Visual interpretations

Keep expanding

This will evolve into a mini learning hub for:

• Market segmentation
• Anomaly detection
• Density estimation
• Understanding manifold structure
• Using PCA/UMAP for visualization
• Advanced clustering models (GMM, Spectral, HDBSCAN)

---

Future Additions

I will extend this repository with subdirectories like:

future-projects/
│
├── Retail RFM Analysis/
├── Customer Loyalty Patterns/
├── Anomaly Detection in Transactions/
├── Product Affinity Clustering/
└── dimensionality-reduction-experiments/

Each with structured README + documentation.

---

Contributing to Myself (or Others)

This repo is primarily a self-learning research space, but structured so others can follow the journey.

You’re welcome to:

• suggest new clustering techniques
• propose datasets
• help reorganize project folders
• or fork and experiment

---