Bank customer churn prediction

The analysis starts by exploring the customer churn dataset: demographics, account details and, most importantly, a churn status for each customer. Visualisations show the distribution of:

• Customer age
• Gender
• Number of dependents
• Education level
• Marital status
• Income category
• Card category
• Months as a customer
• Total number of accounts held
• Months of inactivity in the past year
• Credit limit
• Total transaction amount
• The proportion of churning vs. non-churning customers

• One-hot encoding. Categorical features such as education level are one-hot encoded.
• SMOTE. Synthetic Minority Oversampling Technique addresses the imbalanced classes, since churners are a smaller group.
• PCA. Principal Component Analysis reduces the dimensionality created by one-hot encoding.
• Models. Random Forest and Support Vector Machine (SVM) classifiers, assessed with 5-fold cross-validation.

The F1 score is the primary metric for comparing the two models.

• Python
• One-hot encoding
• SMOTE
• PCA
• Random Forest
• SVM
• 5-fold cross-validation

Authors: Bohdan Boiprav & Catalin Bondari. Data: Kaggle BankChurners. Licensed under the MIT License. Code on GitHub.