AI Dictionary › Fondamenti AI

ROC curve and AUC

Curva ROC e AUC

The ROC curve is a plot showing how a binary classifier's performance changes as the decision threshold used to separate positive from negative cases varies, relating the model's ability to identify true positives to its tendency to generate false alarms. AUC, the area under that curve, summarizes the whole plot in a single number between 0 and 1, indicating how good the model is at distinguishing the two classes regardless of the chosen threshold.

Definition

How it works

A classification model does not directly produce a label, but a probability or score, and a threshold is needed to turn it into a binary decision. The ROC curve plots, for every possible threshold, the true positive rate on the vertical axis and the false positive rate on the horizontal axis: a perfect model touches the top-left corner, a model guessing at random traces a diagonal line, and a good model gets as close as possible to the corner. An AUC of 0.5 indicates performance equivalent to a coin flip, an AUC of 1 indicates perfect separation between classes.

Applications

It is especially useful because it does not require fixing a specific threshold in advance, and it allows different models to be compared on their intrinsic ability to discriminate between classes, regardless of how they will later be calibrated in production. It is widely used in medical diagnostics, where the method's origins trace back to radar signal analysis during World War II, in fraud detection, and in any scoring system where the trade-off between sensitivity and false alarms must be set.

History & etymology

The acronym ROC stands for receiver operating characteristic, an expression born in radar signal processing during the 1940s, used to evaluate how well a radar operator could distinguish a real enemy target from background noise. The method migrated to experimental psychology and medical diagnostics in the 1960s and 70s, before becoming an established standard for evaluating machine learning classifiers starting in the 1980s and 90s.

Related terms

• Precision and Recall
• Confusion matrix
• Classification
• Evaluation metrics
• Imbalanced data

More in Fondamenti AI

• LLM (Large Language Model)
• Token
• Context Window
• Temperature
• AI Hallucination
• Fine-tuning
• RAG (Retrieval Augmented Generation)
• Embedding

Put it into practice

• Real scenario: apply this concept with an instant AI verdict.
• Daily gym: short drills to train dictionary concepts.
• Playoff game: challenge other players, elimination-style.

📱 Download the Android app (beta) iOS coming soon

Say what you mean. Get what you need.

Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.