Precision-Recall Curve & Average Precision Calculator

One point per distinct score. Sort rows by score from high to low. For each distinct score t, predict positive for every row with score at least t, then count true positives (TP) and false positives (FP). Precision = TP / (TP + FP) and recall = TP / (all positives). Tied scores are one threshold, so three rows at 0.6 move the curve in a single step, exactly as scikit-learn precision_recall_curve groups them.

Worked example. The scikit-learn docs use labels [0, 0, 1, 1] and scores [0.1, 0.4, 0.35, 0.8]. Thresholds 0.8, 0.4, 0.35 and 0.1 give precision 1, 0.5, 0.667, 0.5 and recall 0.5, 0.5, 1, 1. The curve then ends at recall 0 with precision 1, a point with no threshold that anchors the plot on the y axis.

Best F1. F1 = 2PR / (P + R) at each threshold. The highlighted row is the threshold with the highest F1, a common starting point when you must pick one cutoff.

Rare positives. ROC curves use the false positive rate, which stays small when negatives vastly outnumber positives, so a model can look excellent on ROC while most of its alerts are false. The PR curve shows that directly through precision. Saito and Rehmsmeier (2015) recommend PR plots for imbalanced binary classification.

Retrieval and detection. AP is the standard summary in information retrieval and object detection, where mean AP (mAP) averages AP across queries or classes. Note that COCO and Pascal VOC use interpolated variants with 101 or 11 recall points, so their numbers differ slightly from scikit-learn's step AP.

Choosing a threshold. Read the table from the top: high thresholds give high precision and low recall. If a false positive costs more than a missed positive, pick a threshold left of the best-F1 point; if misses cost more, move right toward higher recall.