ROC, AUC และค่าเกณฑ์

Most classifiers do not just say yes or no. They give a probability, like 0.73, and you decide where to draw the line. 🎚️

A threshold turns a probability into a label. Above it means positive, below means negative. The default 0.5 is just one choice.

Lower the threshold and you catch more positives but raise false alarms. Raise it and you are stricter but miss more. Each setting trades off.

As the threshold slides, watch the true positive rate against the false positive rate. Together they describe the model at every cutoff.

The ROC curve plots true positive rate versus false positive rate across all thresholds, showing the full trade-off in one picture.

A curve that hugs the top-left corner is excellent: high catch rate with few false alarms. A diagonal line means random guessing.

AUC squeezes the whole ROC curve into one number: the area beneath it. It summarizes performance across every threshold at once.

AUC of 1.0 is perfect, 0.5 is coin-flip random. Read it as the chance the model ranks a true positive above a true negative.

Because AUC ignores any single cutoff, it is great for comparing models. The winner ranks positives ahead of negatives more often.

AUC picks the model, but you still set the threshold for real use, tuning it to favor recall or precision as the problem demands.

On very imbalanced data, ROC can look rosy. A precision-recall curve often tells the harder truth about rare-class performance.

Let's confirm what AUC really tells you about a classifier.

The threshold turns probabilities into labels, the ROC curve shows every trade-off, and AUC scores the whole curve at once. 🎯