Two reports evaluate the same binary classifier on the same documents, yet their precision figures differ. Before looking for a software fault, check which outcome each report treats as positive.
The scikit-learn precision-recall guide defines these measures in terms of the positive class. “Positive” is an evaluation label, not a synonym for a desirable business outcome.
Turn the same fictional table around
Imagine 100 documents. Twenty genuinely need review and eighty do not. The classifier flags ten: eight need review and two do not. It leaves ninety unflagged: twelve still need review and seventy-eight do not.
With “needs review” as positive, precision is 8/10, or 80 percent, and recall is 8/20, or 40 percent. Now make “does not need review” the positive class without changing a single prediction.
The unflagged documents become positive predictions. Of those ninety, seventy-eight are correct, so precision becomes 78/90, approximately 86.7 percent. Of the eighty documents that genuinely do not need review, seventy-eight were identified: recall is 97.5 percent.
That large recall figure does not mean the model has suddenly found more review-needed documents. The question has changed. The twelve missed review cases remain exactly where they were.
Carry the label definition into the export
Write the positive label beside every result, along with the dataset and underlying counts. Check the software’s label encoding and positive-label parameter before comparing reports produced by different teams. A numeric one in a file may mean a different outcome from the one in a presentation.
For a quick validation exercise, calculate both versions from one small confusion table and compare them with the exported metrics. If the values do not reconcile, inspect the label order before changing the model.
Reporting both classes can reveal useful differences. Presenting one class’s strong recall as evidence about the other creates a misleading story from otherwise correct arithmetic.