Myntra Customer Segmentation — Unsupervised ML

RFM-based customer segmentation using K-Means and Agglomerative Hierarchical Clustering for Myntra Gifts Ltd. — identifying Champions, Loyal, New, and At-Risk customers to enable targeted marketing strategies.

---

Table of Contents

• Business Problem
• Dataset Overview
• Methodology
• EDA Highlights
• Customer Segments
• Model Results
• Project Structure
• How to Run
• Tech Stack

---

Business Problem

Myntra Gifts Ltd. is a UK-based online retailer specialising in unique all-occasion giftware. The company operates exclusively online and serves customers across 38 countries.

The Problem: All customers were treated uniformly — same marketing messages, same promotions, same communication strategy regardless of purchase behaviour. A customer who buys every month and spends thousands received identical treatment to one who purchased once two years ago.

The Goal: Apply unsupervised machine learning to segment the customer base into distinct behavioural groups using RFM (Recency, Frequency, Monetary) analysis, then assign targeted business strategies to each segment to:

• Improve customer retention
• Reduce churn among at-risk customers
• Maximise revenue from high-value segments
• Allocate marketing budget efficiently

---

Dataset Overview

Property	Value
Source	Myntra Gifts Ltd. — Online Retail Transactions
Period	December 2009 – December 2011
Raw Rows	541,909
Columns	8
Countries	38
Unique Customers (after cleaning)	~4,300

Dataset URL (Google Sheets):

https://docs.google.com/spreadsheets/d/1_W3Jfp1bTWpPFmqyGgGXYJGd0rHIV8dD/export?format=csv&gid=501524341

Columns:

Column	Type	Description
InvoiceNo	Object	Unique transaction ID. Values starting with 'C' = cancellation
StockCode	Object	Unique product identifier
Description	Object	Product name
Quantity	Integer	Units purchased. Negative = returns
InvoiceDate	Object → Datetime	Transaction timestamp
UnitPrice	Float	Price per unit in GBP
CustomerID	Float → Integer	Unique customer identifier (24.93% missing)
Country	Object	Customer's country

---

Methodology

Raw Data (541,909 rows)
        │
        ▼
  Data Wrangling
  ├── Convert InvoiceDate → datetime
  ├── Drop missing CustomerID (~135,080 rows)
  ├── Remove cancelled transactions (InvoiceNo starts with 'C')
  ├── Remove Quantity ≤ 0 and UnitPrice ≤ 0
  ├── Engineer TotalAmount = Quantity × UnitPrice
  └── Remove duplicate rows
        │
        ▼
  Exploratory Data Analysis
  └── 13 charts: univariate + bivariate + temporal + RFM distributions
        │
        ▼
  Hypothesis Testing (3 tests)
  ├── H1: UK vs non-UK order values (Mann-Whitney U)
  ├── H2: Frequency vs Monetary correlation (Spearman)
  └── H3: Q4 vs non-Q4 revenue (Mann-Whitney U, one-sided)
        │
        ▼
  Feature Engineering → RFM Table
  ├── Recency  : days since last purchase
  ├── Frequency: unique invoices per customer
  └── Monetary : total GBP spend per customer
        │
        ▼
  Preprocessing
  ├── Outlier capping at 99th percentile (Winsorization)
  ├── StandardScaler (zero mean, unit variance)
  └── PCA to 2 components (for visualisation only)
        │
        ▼
  ML Models
  ├── K-Means Clustering (Elbow + Silhouette → K=4)
  └── Agglomerative Hierarchical Clustering (Ward linkage, K=4)
        │
        ▼
  Cluster Profiling → 4 Business Segments
  ├── Champions
  ├── Loyal Customers
  ├── New Customers
  └── At-Risk Customers
        │
        ▼
  Deployment
  ├── kmeans_customer_segmentation.pkl
  └── rfm_standard_scaler.pkl

---

EDA Highlights

Key findings across 13 charts:

• UK Dominance: The United Kingdom accounts for over 90% of all transactions — the business is heavily UK-centric, presenting both a focus opportunity and a concentration risk.
• Strong Seasonality: Revenue peaks every November–December (holiday gifting season). Year-over-year growth from 2010 to 2011 confirms business expansion.
• B2B Buying Pattern: Orders peak between 10 AM–2 PM on Wednesday and Thursday — consistent with trade/wholesale buyers ordering during UK business hours.
• Right-Skewed Spend: All RFM features are right-skewed. A small proportion of customers drive a disproportionate share of revenue — confirming the Pareto principle in this dataset.
• Low Cancellation Rate: ~2% of transactions are cancellations — small in volume but representing real operational cost.
• Price Distribution: Most products are priced GBP 1–5, making the business volume-dependent.

Hypothesis Testing Results:

Hypothesis	Test	Result
UK vs non-UK average order value differs	Mann-Whitney U (two-sided)	Reject H₀ — significant difference
Frequency & Monetary are positively correlated	Spearman rank correlation	Reject H₀ — strong positive correlation
Q4 revenue is higher than non-Q4	Mann-Whitney U (one-sided)	Reject H₀ — Q4 significantly higher

---

Customer Segments

Four segments were identified using K-Means clustering on standardised RFM features.

Segment	Recency	Frequency	Monetary	Size	Business Strategy
🏆 Champions	Very Low	Very High	Very High	~15%	VIP loyalty rewards, early product access, referral programmes
💛 Loyal Customers	Low–Medium	Medium–High	Medium–High	~25%	Upsell premium products, personalised recommendations, loyalty points
🆕 New Customers	Low	Low	Low	~30%	Welcome email series, onboarding discount, showcase bestsellers
⚠️ At-Risk Customers	Very High	Low	Low	~30%	Win-back campaign with heavy discount, satisfaction survey, retargeted ads

RFM Profile Interpretation:

• Recency: Lower = purchased more recently (better)
• Frequency: Higher = orders more often (better)
• Monetary: Higher = spends more in total (better)

---

Model Results

Model	Algorithm	K	Silhouette Score
Model 1	K-Means Clustering	4	~0.38
Model 2	Agglomerative Clustering (Ward)	4	~0.35

Final Model: K-Means with K=4

Selected because:

1. Higher Silhouette Score than Agglomerative Clustering
2. Computationally scalable to full dataset (4,300+ customers)
3. Cluster centroids can be saved and reused to score new customers
4. Both algorithms independently agreed on K=4, validating the 4-segment structure

Hyperparameter Tuning:

• K-Means: Elbow Method + Silhouette Score sweep across K = 2 to 10 → optimal at K=4
• Agglomerative: Ward vs Complete vs Average linkage comparison → Ward wins

---

Project Structure

myntra-customer-segmentation/
├── notebooks/
│   └── Myntra_Customer_Segmentation.ipynb   # Full analysis notebook
├── data/
│   └── README_data.md                        # Dataset source & description
├── models/
│   ├── kmeans_customer_segmentation.pkl      # Trained K-Means model
│   └── rfm_standard_scaler.pkl               # Fitted StandardScaler
├── outputs/
│   ├── rfm_segments.csv                      # CustomerID + RFM + Segment label
│   └── charts/                               # All 13 EDA charts as PNG
├── reports/
│   └── Myntra_Segmentation_Report.pdf        # Project summary report
├── .gitignore
├── requirements.txt
└── README.md

---

How to Run

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/myntra-customer-segmentation.git
cd myntra-customer-segmentation

2. Install dependencies

pip install -r requirements.txt

3. Launch the notebook

jupyter notebook notebooks/Myntra_Customer_Segmentation.ipynb

4. Run all cells — the notebook loads data directly from the Google Sheets URL. No CSV download needed.

5. Predict a new customer's segment

import joblib
import pandas as pd

model  = joblib.load('models/kmeans_customer_segmentation.pkl')
scaler = joblib.load('models/rfm_standard_scaler.pkl')

new_customer = pd.DataFrame({
    'Recency':   [10],   # days since last purchase
    'Frequency': [8],    # number of orders
    'Monetary':  [450]   # total spend in GBP
})

scaled    = scaler.transform(new_customer)
cluster   = model.predict(scaled)
print(f'Predicted cluster: {cluster[0]}')  # → Champions

---

Tech Stack

Library	Purpose
pandas	Data loading, wrangling, aggregation
numpy	Numerical operations
matplotlib / seaborn	All 13 EDA charts
scikit-learn	StandardScaler, KMeans, AgglomerativeClustering, PCA, silhouette_score
scipy	Hierarchical linkage, dendrogram, Mann-Whitney U, Spearman
joblib	Model serialisation and loading

---

Author

[Your Name] Capstone Project — Unsupervised Machine Learning AlmaBetter / [Your Institute Name]

---

License

This project is licensed under the MIT License.