Skip to content
ML Academy

Introduced on day 13

Cost matrix

A table of the cost (or value) of each combination of predicted action and true outcome, from which a decision threshold is derived.

In detail

For a binary flag the four cells are false positive, false negative, true positive and true negative. With zero cost for correct calls, a calibrated probability should trigger action when it exceeds C_FP / (C_FP + C_FN). The matrix needs business inputs (cost of a review, value of a recovered seat) and a stated range; the threshold should be shown with its sensitivity to them.