Evaluate Classification Models

Read the confusion matrix

from sklearn.metrics import classification_report, confusion_matrix
print(confusion_matrix(y_test, preds))
print(classification_report(y_test, preds))

Which metric?

• Precision — trust of positive predictions (spam filters).
• Recall — coverage of actual positives (disease screening).
• F1 — balance of the two.

Threshold tuning

Classifiers output probabilities; the 0.5 cutoff is arbitrary. Adjusting it trades precision for recall — pick the point that matches the real-world cost of each error type.