Customer Behavior Prediction - E-commerce Dataset Churn Predictor

Project Overview

This project demonstrates a machine learning classification model built with real-world e-commerce customer data from Kaggle. The model predicts customer purchase intent using behavioral and demographic features, achieving 82% accuracy.

Quick Start

This repository contains the Customer Behavior Prediction project that was previously part of the data-science-analytics-projects collection.

Dataset Information

• Source: Kaggle E-commerce Dataset
• Total Records: 12,330 customer session events
• Features: 18 behavioral and demographic variables
• Target Variable: Purchase (Binary: 0=No Purchase, 1=Purchase)
• Features:
  • Administrative Pages Visited
  • Informational Pages Visited
  • Product Pages Visited
  • Bounce Rate
  • Exit Rate
  • Page Values
  • Special Day
  • Month
  • Operating System
  • Browser Type
  • Region
  • Traffic Type
  • Visitor Type
  • Weekend (Boolean)
  • Session Duration

Project Structure

01_Customer_Behavior_Prediction/
├── README.md
├── customer_prediction.ipynb
├── customer_prediction.py
├── requirements.txt
├── data/
│   ├── raw/
│   │   └── customer_data.csv
│   └── processed/
│       └── features_engineered.csv
├── models/
│   └── best_model.pkl
└── results/
    ├── confusion_matrix.png
    ├── roc_curve.png
    └── feature_importance.png

Key Achievements

✅ 82% Accuracy - RandomForest Classification
✅ 0.80 Precision - Low false positives
✅ 0.84 Recall - Good positive detection
✅ 0.88 ROC-AUC - Excellent discrimination
✅ Feature Engineering - RFM scores, session quality metrics
✅ Class Imbalance Handling - SMOTE application

Key Metrics

• Accuracy: 82%
• Precision: 80%
• Recall: 84%
• ROC-AUC: 0.88
• Dataset: 12,330 customer session events
• Features: 18 behavioral variables

Methodology

1. Exploratory Data Analysis (EDA)

• Statistical analysis of customer behavior
• Distribution analysis of features
• Correlation matrix and heatmaps
• Missing value assessment
• Class balance analysis

2. Data Preprocessing

• Handling missing values
• Outlier detection and treatment
• Feature scaling (StandardScaler)
• Categorical encoding (LabelEncoder)

3. Feature Engineering

• RFM Scores: Recency, Frequency, Monetary segments
• Behavioral Metrics: Session quality, engagement score
• Temporal Features: Day of week, seasonality
• Device Features: Browser and OS encoding

4. Model Development

• Train-test split (80-20)
• SMOTE for class imbalance
• Random Forest with GridSearchCV
• Hyperparameter tuning
• Cross-validation (5-fold)

5. Model Evaluation

• Classification metrics
• Confusion matrix analysis
• ROC-AUC curve
• Feature importance analysis

Results & Insights

Model Performance

• Accuracy: 82%
• Precision: 80%
• Recall: 84%
• F1-Score: 82%
• ROC-AUC: 0.88

Top Predictive Features

1. Page Values (importance: 24%)
2. Product Pages Visited (importance: 18%)
3. Session Duration (importance: 16%)
4. Bounce Rate (importance: 14%)
5. Visitor Type (importance: 12%)

Business Insights

• High page values strongly indicate purchase intent
• Product page engagement is critical
• Session duration > 5 minutes correlates with 70% purchase probability
• Weekend visitors have 15% higher purchase rate
• Returning visitors have 35% higher conversion rate

Technologies Used

• Python 3.9+
• Pandas - Data manipulation and analysis
• NumPy - Numerical computations
• Scikit-learn - Machine learning algorithms
• Matplotlib & Seaborn - Data visualization
• Imbalanced-learn - SMOTE for class balance
• Jupyter Notebook - Interactive analysis

How to Run

Prerequisites

pip install -r requirements.txt

Or install manually:

pip install pandas numpy scikit-learn matplotlib seaborn imbalanced-learn jupyter

Execute Analysis

# Option 1: Run Jupyter Notebook
jupyter notebook customer_prediction.ipynb

# Option 2: Run Python script
python customer_prediction.py

Files Description

File	Description
customer_prediction.ipynb	Complete interactive analysis notebook
customer_prediction.py	Reproducible Python script
requirements.txt	All dependencies
data/raw/customer_data.csv	Original Kaggle dataset
data/processed/features_engineered.csv	Processed features
models/best_model.pkl	Trained RandomForest model

Future Improvements

• XGBoost and LightGBM model comparison
• Deep learning neural network approach
• Feature interaction analysis
• Ensemble methods stacking
• Real-time prediction API

📚 Resources & Data Sources

Project 1: Customer Behavior Prediction

• Primary Source: Kaggle E-commerce Dataset
• Dataset Link: UCI Machine Learning Repository - Online Retail II
• Alternative Source: UCI ML Repository
• Description: Real-world e-commerce customer transaction data
• License: CC0: Public Domain

Author

Nagendra Singh Rawat | Data Science & Analytics

References

• Kaggle: E-commerce Dataset
• Scikit-learn Documentation
• SMOTE: Synthetic Minority Over-sampling Technique