🔍 KDD Cup 2009 Customer Relationship Management Prediction

A machine learning project for predicting customer churn, appetency, and upselling using the Orange dataset from the KDD Cup 2009 competition.

📋 Description

This repository contains a comprehensive machine learning solution for customer relationship management (CRM) prediction tasks. It focuses on three key prediction challenges from the KDD Cup 2009 competition:

• Churn: Predicting which customers are likely to leave
• Appetency: Predicting which customers are likely to buy a product
• Upselling: Predicting which customers are likely to buy more products

The project implements multiple machine learning models, preprocessing strategies, and evaluation techniques to achieve high-performance predictions on these tasks.

✨ Features

• Multiple Classifier Support: Implementation of 6 different classifiers:

  • Random Forest (RFC)
  • Decision Tree (DTC)
  • AdaBoost (ABC)
  • Gradient Boosting (GBC)
  • Bagging (BGC)
  • Voting Classifier (VTC)

• Advanced Preprocessing:

  • Three different preprocessing strategies (DS01, DS02, DS03)
  • Handling of missing values with various imputation techniques
  • Feature selection and transformation
  • Categorical feature encoding

• Hyperparameter Tuning:

  • Single parameter grid search
  • Multi-parameter grid search
  • Optimized configurations for each target variable

• Comprehensive Evaluation:

  • ROC AUC scoring
  • Confusion matrix analysis
  • Threshold adjustment for classification
  • Feature importance ranking

• Visualization:

  • Missing data visualization
  • Confusion matrix plots
  • Performance comparison charts
  • Decision tree visualization

🛠️ Setup Guide

Prerequisites

• Python 3.x
• Required Python packages (install via pip):

  pandas
  numpy
  scikit-learn
  matplotlib
  seaborn
  pydotplus
  missingno
  

• Graphviz (for decision tree visualization)

Installation

1. Clone this repository:

   git clone https://github.com/yourusername/KDDCup2009.git
   cd KDDCup2009
   

2. Install required packages:

   pip install -r requirements.txt
   

3. Download the Orange dataset from the KDD Cup 2009 competition and place it in the data directory.

🚀 Usage

The main script accepts various command-line arguments to control the execution flow:

python main.py PrepEnabled=1 ProcessDS01=1 ProcessDS02=1 ProcessDS03=1 PredictChurn=1 PredictAppetency=1 PredictUpselling=1 GridSearchSingleRFC=0 GridSearchSingleDTC=0 GridSearchSingleABC=0 GridSearchSingleGBC=0 GridSearchSingleBGC=0 GridSearchMultiRFC=0 GridSearchMultiDTC=0 GridSearchMultiABC=0 GridSearchMultiGBC=0 GridSearchMultiBGC=0 FinalRFC=1 FinalDTC=1 FinalABC=1 FinalGBC=1 FinalBGC=1 FinalVTC=1 BaselineRFC=1 BaselineDTC=1 BaselineABC=1 BaselineGBC=1 BaselineBGC=1 BaselineVTC=1 PlotGraphs=1

Command-line Arguments

• Data Preparation:

  • PrepEnabled=1: Enable data preprocessing
  • ProcessDS01=1: Process dataset with strategy 1 (mean imputation)
  • ProcessDS02=1: Process dataset with strategy 2 (zero imputation)
  • ProcessDS03=1: Process dataset with strategy 3 (special value imputation + one-hot encoding)

• Prediction Targets:

  • PredictChurn=1: Enable churn prediction
  • PredictAppetency=1: Enable appetency prediction
  • PredictUpselling=1: Enable upselling prediction

• Grid Search:

  • GridSearchSingleXXX=1: Enable single parameter grid search for classifier XXX
  • GridSearchMultiXXX=1: Enable multi-parameter grid search for classifier XXX

• Model Evaluation:

  • BaselineXXX=1: Evaluate baseline model for classifier XXX
  • FinalXXX=1: Evaluate final (optimized) model for classifier XXX

• Visualization:

  • PlotGraphs=1: Generate visualization plots

📊 Project Structure

KDDCup2009/
├── main.py                 # Main entry point
├── preprocessor.py         # Data preprocessing functionality
├── filehandler.py          # File I/O operations
├── modeller.py             # Machine learning models implementation
├── visualizer.py           # Visualization functions
├── logging.conf            # Logging configuration
├── data/                   # Data directory
│   ├── orange_small_train.csv                    # Training data
│   ├── orange_small_test.csv                     # Test data
│   ├── orange_small_train_churn.labels.csv       # Churn labels
│   ├── orange_small_train_appetency.labels.csv   # Appetency labels
│   ├── orange_small_train_upselling.labels.csv   # Upselling labels
│   └── ...                                       # Other data files
└── graphs/                 # Generated visualizations
    ├── Data Completion Categorical Features.png
    ├── Graph - Baseline vs Final - Churn Scores - DS02.png
    └── ...                 # Other visualization files

📈 Results

The project evaluates multiple classifiers on three prediction tasks (churn, appetency, upselling) using ROC AUC as the primary evaluation metric. Results are saved in CSV files in the data directory and visualized in the graphs directory.

Key findings:

• Gradient Boosting and Random Forest classifiers generally perform best
• Feature importance varies significantly between prediction tasks
• Ensemble methods (Voting Classifier) can improve performance by combining multiple models

🔧 Advanced Usage

Custom Preprocessing

You can modify the preprocessing strategies in preprocessor.py to experiment with different approaches:

• Change imputation strategies for missing values
• Adjust feature selection criteria
• Implement additional feature engineering techniques

Model Tuning

To tune a specific model for a particular prediction task:

1. Enable the appropriate grid search parameters
2. Run the main script
3. Analyze the results in the output CSV files
4. Update the final model parameters in modeller.py

📚 Resources

• KDD Cup 2009 Competition
• Orange Dataset Description
• Scikit-learn Documentation

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.