$ CAT CASE-STUDY.MD · SVM · LOGISTIC REGRESSION
Customer churn prediction
SVM vs. logistic regression.
A model that classifies customers as likely to churn or stay, comparing SVM kernels against regularized logistic regression on a Kaggle dataset.
01 · BREAKDOWN
Project breakdown.
Data preprocessing, a train-test split, SVM training and evaluation, then regularized logistic regression. Feature selection and hyperparameter tuning made a large difference.
Models were compared on accuracy, precision, recall, F1, confusion matrices, and ROC curves.
Build a churn prediction model
Accurately identify the customers most likely to leave.
Evaluate SVM kernels
Compare linear, polynomial, and RBF kernels to find the best performer.
Compare against logistic regression
Test whether regularized logistic regression can match the SVMs.
02 · FINDINGS
Key findings.
Polynomial SVM won, at 0.83 AUC.
The polynomial-kernel SVM was the best model for churn. Logistic regression is a simpler, more interpretable alternative with comparable performance.
Polynomial kernel led
Highest recall and AUC, making it the best choice for predicting churn.
Linear kernel fell short
A linear decision boundary couldn't capture the relationships in this data.
Tuning mattered
Tuning C, polynomial degree, and RBF gamma significantly improved accuracy.
