Confusion Matrix Calculator: F1, Precision, Recall, MCC

In Binary mode you type the four cells of a 2 by 2 confusion matrix. Every metric is computed from those counts with the standard scikit-learn definitions, and any ratio whose denominator is 0 is shown as n/a instead of NaN, with a note explaining why.

• Precision = TP / (TP + FP)
• Recall (sensitivity) = TP / (TP + FN)
• Specificity = TN / (TN + FP)
• F1 = 2TP / (2TP + FP + FN)
• F-beta = (1 + b2)TP / ((1 + b2)TP + b2 FN + FP), where b2 is beta squared
• MCC = (TP x TN - FP x FN) / sqrt((TP + FP)(TP + FN)(TN + FP)(TN + FN))

In Multi-class mode you pick 3 to 8 classes, name them, and fill a grid where rows are the actual class and columns the predicted class. You get a per-class precision, recall and F1 report, macro, weighted and micro averages, balanced accuracy, Cohen's kappa and Gorodkin's multi-class MCC.

Pick the metric that matches the cost of each error. When a missed positive is expensive, such as a disease screen or fraud alert, raise beta above 1 so F-beta weights recall more. When false alarms are expensive, use beta below 1.

• Report MCC alongside F1. Chicco and Jurman (2020) argue MCC is more informative because it only scores high when both the positive and negative classes are predicted well.
• F1 ignores true negatives entirely, so adding 10,000 more correct negatives leaves it unchanged.
• Use balanced accuracy or macro F1 for multi-class sets where one class dominates.
• Kappa and MCC range from -1 to 1, where 0 means no better than chance.