Customer Segmentation & Recommendation System for Retail Stores

K-Means customer segmentation on 200 mall customers — three clustering experiments across age, income, and spending dimensions, with 3D interactive visualisation. Segments directly map to targeted product recommendations.

Overview · Dataset · Segmentation Results · How Segments Map to Recommendations · Visualisations · Setup · Code Fixes

---

Overview

Retail stores lose revenue by treating every customer the same. A customer who earns $120k/year and spends freely needs completely different recommendations than one who earns $20k and shops carefully. This project uses K-Means clustering to segment 200 mall customers into distinct behavioural groups — each segment then maps directly to a targeted recommendation strategy.

The project runs three independent clustering experiments, progressively adding dimensions to discover richer customer patterns:

Experiment	Features used	Optimal K	Business insight
1	Age × Spending Score	4 clusters	Young high-spenders vs cautious older customers
2	Annual Income × Spending Score	5 clusters	The classic 5-segment retail model
3	Age × Annual Income × Spending Score	6 clusters	Full 3D behavioural profiling (interactive Plotly)

---

Dataset

Mall Customers Dataset — 200 retail mall customers

Feature	Type	Range	Mean
CustomerID	Integer	1 – 200	—
Gender	Categorical	Male / Female	112F / 88M
Age	Integer	18 – 70	38.8 years
Annual Income (k$)	Integer	$15k – $137k	$60.6k
Spending Score (1-100)	Integer	1 – 99	50.2

200 customers · 5 features · 0 missing values · 112 Female / 88 Male
Age range: 18–70 · Income range: $15k–$137k · Spending score: 1–99

Key correlations found during EDA:

• Annual Income and Spending Score have near-zero correlation overall — but strong within clusters
• Young customers (18–30) show bimodal spending: either very high OR very low
• Female customers slightly outspend male customers at equivalent income levels

---

Segmentation Results

Experiment 1 — Age × Spending Score (K=4)

Elbow method identifies 4 optimal clusters:

Cluster  │  Age range  │  Spending Score  │  Segment name
─────────┼─────────────┼──────────────────┼──────────────────────
  0      │  20 – 40    │     60 – 99      │  Young High Spenders
  1      │  40 – 70    │     40 – 60      │  Mature Moderate Spenders
  2      │  18 – 35    │      1 – 40      │  Young Cautious Buyers
  3      │  30 – 65    │     60 – 99      │  Mid-Age High Spenders

Experiment 2 — Annual Income × Spending Score (K=5)

This is the most business-relevant segmentation — elbow identifies 5 classic retail segments:

Cluster  │  Income     │  Spending Score  │  Segment name              │  Size
─────────┼─────────────┼──────────────────┼────────────────────────────┼──────
  0      │  Low        │  Low (1–40)      │  Careful Customers          │  ~35
  1      │  Low        │  High (60–99)    │  Impulsive Customers        │  ~22
  2      │  Medium     │  Medium (40–60)  │  Standard Customers         │  ~78
  3      │  High       │  Low (1–40)      │  Conservative High-Earners  │  ~22
  4      │  High       │  High (60–99)    │  Target Customers ⭐        │  ~23

Cluster 4 (Target Customers) = the highest-value retail segment. High income + high willingness to spend. These customers should receive premium product recommendations and loyalty benefits.

Experiment 3 — Age × Income × Spending Score (K=6, 3D)

Three-dimensional segmentation reveals 6 distinct clusters visualised in an interactive 3D Plotly scatter plot. The additional age dimension splits the "Standard Customers" group into younger and older sub-segments, enabling age-appropriate recommendations on top of income-based ones.

---

How Segments Map to Recommendations

The clusters aren't just interesting patterns — they directly drive a recommendation strategy:

Customer arrives at store
         │
         ▼
┌─────────────────────┐
│  Predict segment    │  Input: Age, Income, Spending Score
│  (trained KMeans)   │  Output: Cluster ID (0–4)
└────────┬────────────┘
         │
         ▼
┌────────────────────────────────────────────────────────────────┐
│                  RECOMMENDATION STRATEGY BY SEGMENT            │
├──────────────────────┬─────────────────────────────────────────┤
│  Cluster 0 (Careful) │  Budget products, value packs, discounts│
│  Cluster 1 (Impulse) │  Flash deals, limited stock alerts      │
│  Cluster 2 (Standard)│  Popular items, bestsellers             │
│  Cluster 3 (Conserv.)│  Premium brands, quality over quantity  │
│  Cluster 4 (Target ⭐)│  Luxury items, premium bundles         │
└──────────────────────┴─────────────────────────────────────────┘

Predicting a new customer's segment:

import numpy as np
from sklearn.cluster import KMeans
import joblib

# Load trained model
model = joblib.load('kmeans_income_spending.pkl')

# New customer: Annual Income = $85k, Spending Score = 72
new_customer = np.array([[85, 72]])
cluster = model.predict(new_customer)[0]

segment_names = {
    0: "Careful Customer",
    1: "Impulsive Customer",
    2: "Standard Customer",
    3: "Conservative High-Earner",
    4: "Target Customer"
}
print(f"Segment: {segment_names[cluster]}")
# Output: Segment: Target Customer

---

Visualisations

The notebook generates 10 visualisation types: 5-cluster Income × Spending plot

Plot	What it shows
Distribution plots (3 features)	Age, Income, Spending Score distributions
Gender count plot	112 Female / 88 Male breakdown
3×3 regression grid	Pairwise relationships between all numeric features
Gender scatter: Age vs Income	Income patterns by age, split by gender
Gender scatter: Income vs Spending	Spending behaviour by income, split by gender
Violin + swarm plots	Feature distributions by gender with individual points
Elbow curve (Experiment 1)	K=1–10 inertia to select K=4
Cluster boundary map (Experiment 1)	Decision regions: Age × Spending
Elbow curve + cluster map (Experiment 2)	K=5 selection + Income × Spending clusters
3D Plotly scatter (Experiment 3)	Interactive 3D: rotate/zoom Age × Income × Spending

---

Project Structure

Recommendation-System-for-Retail-Stores/
│
├── RecomndatonSysRetailStore.ipynb    ← Main Jupyter notebook (41 cells)
├── Mall_Customers.csv                  ← Dataset: 200 customers, 5 features
├── requirements.txt                    ← Python dependencies
├── .gitignore
└── README.md

---

Setup

Run in Google Colab (recommended — zero setup)

Click the badge at the top of this README:

Then upload Mall_Customers.csv when prompted, or mount Google Drive.

Run locally

# 1. Clone the repo
git clone https://github.com/imAryanSingh/Recommendation-System-for-Retail-Stores.git
cd Recommendation-System-for-Retail-Stores

# 2. Install dependencies
pip install -r requirements.txt

# 3. Fix the dataset path in Cell 4 (see Code Fixes below)

# 4. Launch
jupyter notebook RecomndatonSysRetailStore.ipynb

requirements.txt

numpy>=1.21.0
pandas>=1.3.0
matplotlib>=3.4.0
seaborn>=0.12.0
scikit-learn>=1.0.0
plotly>=5.0.0
jupyter>=1.0.0

---

Code Fixes

Three lines in the notebook need updating for modern library versions and local use:

Fix 1 — Dataset path (Cell 4)

# CURRENT (Colab-only — breaks locally):
dataset = pd.read_csv('/content/Mall_Customers.csv')

# FIX (works everywhere):
import os
dataset = pd.read_csv('Mall_Customers.csv')   # if running from repo root

Fix 2 — Deprecated sns.distplot (Cell 17)

sns.distplot was removed in Seaborn 0.12+. Replace with sns.histplot:

# CURRENT (raises FutureWarning / error in seaborn >= 0.12):
sns.distplot(dataset[x], bins=20)

# FIX:
sns.histplot(dataset[x], bins=20, kde=True)

Fix 3 — Deprecated KMeans algorithm='elkan' (Cells 27, 33, 39)

# CURRENT (deprecated in scikit-learn 1.1+):
algorithm = KMeans(n_clusters=4, ..., algorithm='elkan')

# FIX:
algorithm = KMeans(n_clusters=4, ..., algorithm='lloyd')

---

Key Concepts Explained

Why K-Means for customer segmentation? K-Means partitions customers into K groups where each customer belongs to the cluster with the nearest centroid. It's fast (O(n·k·i)), interpretable, and produces actionable segments — unlike deep learning models, you can directly describe what each cluster means in business terms.

Why the Elbow Method? Inertia (sum of squared distances from each point to its cluster centroid) always decreases as K increases — but the rate of decrease sharply slows after the optimal K. The "elbow" in the inertia curve marks this point. Experiment 1 → K=4, Experiment 2 → K=5, Experiment 3 → K=6.

Why k-means++ initialisation? Random centroid initialisation often leads to poor convergence (local minima). init='k-means++' spreads initial centroids far apart, giving much more consistent cluster quality across random seeds.

---

Technologies

Library	Version	Purpose
Pandas	≥1.3	Data loading, EDA, describe/info/corr
NumPy	≥1.21	Array operations, meshgrid for decision boundaries
Scikit-learn	≥1.0	KMeans clustering, elbow computation
Matplotlib	≥3.4	All 2D plots, cluster boundary visualisation
Seaborn	≥0.12	Distplots, violin plots, swarm plots, regplots
Plotly	≥5.0	Interactive 3D scatter for 3-feature clustering

---

About the Author

Aryan Singh — AI/ML Engineer

Developed during AI Vicharana Shala residential programme at IIT Ropar (May–Jul 2024) B.Tech CSE · Mohanlal Sukhadia University · GATE 2026 (88.31 percentile)

---

Also see

• Wake-Word Detection — ISRO TRISHNA Satellite
• Wildfire Prediction from Satellite Imagery
• Smart Vision Quality Control — Top 0.3% Flipkart GRID
• Image · Audio · Emoji Steganography
• E-Commerce Sales Dashboard