Choisir le seuil et la métrique

Classifiers output a probability, then a threshold turns it into a label. The default 0.5 is rarely best for rare classes.

Lowering the threshold flags more items as the rare class, catching more true cases at the cost of extra false alarms.

To tune a cutoff you need scores, not just labels. Ask the model for the predicted probability of the positive class.

Decide what hurts more. Missing rare cases means optimize recall; too many false alarms means favor precision.

When you cannot pick a side, the F1 score blends precision and recall into one number worth maximizing.

For skewed data the precision-recall curve tells the real story far better than an ROC curve does.

Try many cutoffs and watch how the trade-off shifts. The PR curve shows every precision-recall pair at once.

Once you pick a value, compare each probability to it. Above the line becomes the positive class, below becomes negative.

ROC AUC can look great on heavy imbalance even when precision is poor, so do not trust it alone here.

Choose your threshold on a held-out validation set, never on the test set, or your reported numbers will lie.

The right metric reflects real-world cost. Let the business goal pick precision, recall, or F1, then tune to it. 🎯

Test your grip on thresholds.

Tune the threshold on validation data, prefer precision, recall, or F1 over accuracy, and trust the PR curve on skew. ✅