EXPERIMENT 01 / TITANIC BASELINE
A score is a start.
Inspect the decisions.
A leakage-safe logistic baseline on the repository’s labeled Titanic data. Explore actual held-out predictions, threshold tradeoffs, and the limits of a small evaluation.
Loading the local experiment report…
01 / CLASSIFICATION TRADEOFF
What the threshold changes
Actual outcome is the row; prediction is the column. Positive means recorded survival. Changing a threshold explores tradeoffs on this holdout; it does not create a new unbiased evaluation.
02 / RANKING QUALITY
ROC across all thresholds
False-positive rate (horizontal) vs true-positive rate (vertical). AUC is undefined when a cohort contains only one actual class. The dashed line is a random-ranking reference.
03 / HELD-OUT EVIDENCE
Inspect individual predictions
| Passenger ID | Recorded sex | Class | Age | Model probability | Actual | Predicted |
|---|
04 / FITTED MODEL
Coefficients, not causes
Coefficients are log-odds weights after training-only preprocessing. Numeric features are standardized; categorical terms are one-hot columns. Correlated inputs and encoding affect their values. These are not causal effects.
05 / REPRODUCTION RECORD
Trace the result to its inputs.
The competition test file has no verified labels here. The repository’s answer.csv is excluded from this benchmark; its provenance does not establish ground truth. No leaderboard score is claimed.
Inspect the experiment pipeline ↗Three projects.
Better experimental habits.
The MNIST notebook now shares one axis-1 L2 preprocessing function for training, handwritten images and CSV inference. The traffic analysis counts crash, vehicle and person records at their own grain, preventing many-to-many joins from multiplying injury totals.
The Titanic result is a single small, passenger-level holdout from a historical educational dataset. Related passengers may span splits; confidence intervals, family-grouped evaluation and broader validation are future work. Demographic slices show sample behavior, not policy recommendations or general population claims.