🧠 Customer Segmentation using K-Nearest Neighbors (KNN)

This project uses the K-Nearest Neighbors (KNN) algorithm to classify customers into segments based on their demographics and spending behavior. We go through building a base model, performing hyperparameter tuning, and analyzing the results with a focus on model performance and generalization.

---

📌 Problem Statement

We aim to predict a customer’s spending behavior category (Low, Medium, High) based on:

• Gender
• Age
• Annual Income

We transformed this into a classification problem by segmenting the Spending Score into 3 categories:

• 0: Low spender
• 1: Medium spender
• 2: High spender

---

📊 Dataset

Mall Customer Segmentation Data
📂 Download from Kaggle: Click here

Features:

• Gender (converted to 0/1)
• Age
• Annual Income (k$)
• Spending Score (1–100) ➜ Used to create target class: Segment

---

✅ Project Workflow

1. Data Cleaning & Preprocessing

   • Encoded categorical variables
   • Normalized features using StandardScaler
   • Created a Segment target variable by binning Spending Score

2. Base KNN Model

   • Used default k=5
   • Achieved high test accuracy: 82.5%
   • However, this could be overfitting to the current train-test split

3. Hyperparameter Tuning (GridSearchCV)

   • Tuned n_neighbors, weights, and distance metric
   • Best parameters: e.g., k=4, weights='uniform', metric='euclidean'
   • Final test accuracy: 77.5%
   • Accuracy slightly dropped but model generalizes better on unseen data

4. Elbow Method Visualization

   • Confirmed k=4 is near the optimal elbow point

---

📈 Results & Interpretation

Model	Accuracy	Comment
Base KNN	82.5%	High accuracy but overfitting
Tuned KNN (CV-based)	77.5%	More stable, better generalization

• The base model may have overfitted the test data (especially due to small dataset size).
• The tuned model, despite lower accuracy, is less biased by specific train-test splits and performs more reliably when predicting new or unseen data.

---

🧪 Libraries Used

• pandas, numpy
• matplotlib, seaborn
• scikit-learn

---

📘 Key Learnings

• Always validate model performance using cross-validation, not just one train-test split.
• Higher accuracy ≠ better model-generalization matters more.
• KNN is sensitive to scaling and the choice of k, so visualization (like Elbow Method) is helpful.

---