Customer Churn Prediction

Project Description

Customer churn prediction on telecommunication data using Random Forest Classifier (AUC=0.904)

churn-prediction.ipynb

An implementation of RandomForestClassifier trained against 3333 observations with 17 features using 10-fold cross validation.
Missing values are imputed by using median (numeric data) or mode (categorical data). Feature vectors are standardized before fitting into the model.
Grid search is used to find the best parameter set of the RandomForestClassifier: {'max_depth': 10, 'n_estimators': 100}.
Evaluation on area under ROC curve is made using 10-fold cross validation. Best AUC score = 0.904 (+/-0.090) for the parameter set {'max_depth': 10, 'n_estimators': 100}.

model-comparison.ipynb

The dataset is also tested against other classifiers, i.e. KNeighborsClassifier, SVC, LinearSVC, SGDClassifier. Grid search is implemented to find the best parameter set of the classifiers on 10-fold CV.
RandomForestClassifier achieves the best AUC score compared with the above classifiers.

churn.csv
churn-clean.csv

Original and cleaned churn dataset of 3333 observations with 17 columns of selected features and the final clolumn of churn labels.