Pondération des classes et seuils

You can fight imbalance without touching the data at all. Two model-side levers help: class weights and the decision threshold.

Class weights tell the model that mistakes on the rare class hurt more. It pays a bigger penalty for missing minority rows.

Many scikit-learn models accept class_weight set to balanced. It auto-weights each class inversely to how often it appears.

Class weights reshape the loss instead of the dataset. No rows are added or dropped, so you keep all your original data.

Most classifiers predict a probability, then label it positive if it tops 0.5. That default cutoff is a choice, not a law.

Lower the threshold and the model flags positives more eagerly. That catches more rare cases, at the price of more false alarms.

To tune a threshold, ask the model for probabilities, not hard labels. Then apply your own cutoff to those scores.

A lower threshold lifts recall but drops precision. Raising it does the reverse. The right point depends on which mistake costs more.

If a missed fraud is far worse than a false alarm, lean toward a lower threshold so the rare class is rarely missed.

Class weights and threshold tuning stack nicely. Weight the rare class during training, then pick a cutoff that matches your real costs.

Choose the threshold using validation data, never the test set. Then apply that same fixed cutoff for an honest final score.

What happens when you lower the decision threshold?

Without resampling, class weights penalize rare-class errors and a tuned threshold trades recall against precision to fit your costs. 🎯