[FAQ Bot] NEW: Why use ROC AUC instead of accuracy?

_questions/machine-learning-zoomcamp/module-4/031_edec02e3a0_why-use-roc-auc-instead-of-accuracy.md

@@ -0,0 +1,18 @@
+---
+id: edec02e3a0
+question: Why use ROC AUC instead of accuracy?
+sort_order: 31
+---
+
+Accuracy depends on a chosen decision threshold (e.g., classify as positive when the predicted probability ≥ 0.5). If you change that threshold, accuracy changes too.
+
+ROC AUC instead measures how well your model ranks positives above negatives across *all* possible thresholds. It’s threshold-independent.
+
+Accuracy can also be misleading with imbalanced classes. For instance:
+- If 58% of examples are positive, a dumb model that predicts “positive” for everyone gets ~58% accuracy without learning anything.
+- If only 1% are positive, predicting “negative” for everyone yields ~99% accuracy while finding none of the positives.
+ROC AUC isn’t fooled by this: a constant predictor scores around 0.5 (random ranking).
+
+A useful interpretation: AUC is the probability that a randomly chosen positive example receives a higher score than a randomly chosen negative example.
+
+So if your goal is to evaluate the model’s ability to separate classes by ranking (not a specific threshold), ROC AUC is usually a better choice than accuracy.