Customer Churn Prediction

This repository contains a machine learning project focused on predicting customer churn for a telecommunications company. The project is implemented in a Jupyter Notebook (client_charn_prediction.ipynb) and uses a dataset to build and evaluate a predictive model. Below is an overview of the project structure, data, methodology, and results.

Project Overview

The goal of this project is to develop a predictive model that determines the likelihood of customer churn based on various features such as demographic information, service usage, and billing details. The model is trained and evaluated using a dataset provided in CSV format, and the final predictions are submitted to Kaggle for scoring.

Dataset

The dataset is split into three files, downloaded from Google Drive:

• train.csv: Training data with features and the target variable (Churn).
• test.csv: Test data for generating predictions.
• submission.csv: Template for submitting predictions to Kaggle.

Features

The dataset includes the following features, categorized into numerical and categorical:

• Numerical Features:
  • ClientPeriod: Duration of the customer's subscription (in months).
  • MonthlySpending: Monthly payment amount.
  • TotalSpent: Total amount spent by the customer.
• Categorical Features:
  • Sex: Customer's gender (Male/Female).
  • IsSeniorCitizen: Whether the customer is a senior citizen (0/1).
  • HasPartner: Whether the customer has a partner (Yes/No).
  • HasChild: Whether the customer has children (Yes/No).
  • HasPhoneService: Whether the customer has phone service (Yes/No).
  • HasMultiplePhoneNumbers: Whether the customer has multiple phone numbers (Yes/No/No phone service).
  • HasInternetService: Type of internet service (DSL/Fiber optic/No).
  • HasOnlineSecurityService: Whether the customer has online security (Yes/No/No internet service).
  • HasOnlineBackup: Whether the customer has online backup (Yes/No/No internet service).
  • HasDeviceProtection: Whether the customer has device protection (Yes/No/No internet service).
  • HasTechSupportAccess: Whether the customer has tech support (Yes/No/No internet service).
  • HasOnlineTV: Whether the customer has online TV (Yes/No/No internet service).
  • HasMovieSubscription: Whether the customer has a movie subscription (Yes/No/No internet service).
  • HasContractPhone: Contract type (Month-to-month/One year/Two year).
  • IsBillingPaperless: Whether billing is paperless (Yes/No).
  • PaymentMethod: Payment method (Electronic check, Mailed check, Bank transfer, Credit card).

Target Variable

• Churn: Binary variable indicating whether the customer churned (1) or not (0).

Methodology

The project follows a structured approach to data analysis and model building:

1. Data Loading and Exploration:

   • The training and test datasets are loaded using pandas.
   • Initial exploration is performed using data.head() to inspect the data structure.
   • Features are categorized into numerical (num_cols) and categorical (cat_cols) for preprocessing.

2. Data Preprocessing:

   • Basic data cleaning is conducted to ensure data quality (specific cleaning steps are not detailed in the notebook but implied by the "Basic data clean" section).
   • Categorical features are identified for use in the CatBoost model.

3. Model Training:

   • A CatBoostClassifier is used as the primary model, leveraging its ability to handle categorical features natively.
   • Hyperparameter tuning is performed using GridSearchCV with 5-fold cross-validation to optimize the following parameters:
     • iterations: [100, 200, 400]
     • learning_rate: [0.01, 0.1, 0.3]
     • depth: [4, 7]
     • l2_leaf_reg: [3.0, 6.0]
   • A refined grid search is conducted with a narrower parameter range:
     • iterations: [90, 95, 100, 105, 110]
     • learning_rate: [0.09, 0.1, 0.11]
     • depth: [4]
     • l2_leaf_reg: [6.0]
   • The model is trained on a GPU for efficiency.

4. Evaluation:

   • The model's performance is evaluated using the ROC-AUC metric.
   • The best model achieves a ROC-AUC score of 0.8441 on the test set.

5. Prediction and Submission:

   • The best model is used to predict churn probabilities for the test dataset.
   • Predictions are formatted according to the Kaggle submission template and saved as my_submission.csv.

Results

• The final model, after hyperparameter tuning, achieves a ROC-AUC score of 0.8441 on the test set.
• The submission file (my_submission.csv) contains predicted churn probabilities for the test dataset, ready for Kaggle evaluation.

Dependencies

The project relies on the following Python libraries:

• pandas: For data manipulation and loading.
• numpy: For numerical operations.
• matplotlib: For plotting (though no plots are explicitly shown in the notebook).
• gdown: For downloading datasets from Google Drive.
• catboost: For the CatBoostClassifier model.
• scikit-learn: For GridSearchCV and evaluation metrics.

File Structure

• client_charn_prediction.ipynb: Main Jupyter Notebook containing the project code.
• train.csv: Training dataset (downloaded during execution).
• test.csv: Test dataset (downloaded during execution).
• submission.csv: Submission template (downloaded during execution).
• my_submission.csv: Final submission file with predictions.

How to Run

1. Clone this repository.
2. Install the required dependencies: pip install pandas numpy matplotlib gdown catboost scikit-learn.
3. Run the Jupyter Notebook (client_charn_prediction.ipynb) in an environment with Jupyter installed.
4. The notebook will download the datasets, train the model, and generate the submission file.