ML Pipeline

Overview

GitHub Actions orchestrates the whole chain and nothing is run by hand. Preprocessing and the classifier are one scikit-learn Pipeline, so imputation is fitted on training folds only and the holdout score isn't inflated by leakage.

Highlights

• Random forest over engineered features: title, family size, deck, is-alone
• 5-fold CV accuracy 0.827 ± 0.040, holdout ROC-AUC 0.848
• Model published to Hugging Face, image to Docker Hub
• Runner pinned to ubuntu-24.04 so the build environment can't drift