Lire les prédictions du modèle

A model gives you more than a single answer. Learning to read its predictions tells you not just what it chose, but how confident it was.

Calling predict returns the single most likely class for each input. This is the hard label, the model committing to one answer.

Use predict_proba to see the chance assigned to each class. A 0.95 spam score means far more conviction than a borderline 0.51.

Each row of probabilities adds up to exactly 1. The model splits all its belief across the available classes, never more and never less.

Probability columns follow the order in classes_. Check that attribute so you know which column means spam and which means ham.

Internally the model works in logs to avoid tiny numbers underflowing. The predict_log_proba method exposes those raw log scores if you need them.

A high probability means the model is sure, not that it is right. Naive Bayes can be overconfident, so treat its numbers with healthy caution.

You need not flag spam at 0.5. Raising the threshold to 0.8 means fewer false alarms but a few more spam messages slipping through.

Find inputs where the model was confident yet wrong. These errors often reveal missing words or labels that need fixing in your data.

For Naive Bayes you can list which words pushed a message toward spam. That word-level evidence makes the model pleasantly easy to explain. 🔎

Probabilities let you sort by risk, route the unsure cases to a human, and set smart cutoffs. Reading them well turns a model into a real tool.

Which method reveals how confident the model is in each class?

You moved beyond hard labels to read probabilities, check class order, tune thresholds, and explain decisions. That is how you trust a classifier. ✅