Bug: Precision and recall calculations

[jonnygw1]

Description:
I discovered that the Precision and Recall have their logic reversed, the Precision metric is actually calculating Recall, and the Recall metric is calculating Precision.

The Issue:
In the Precision implementation, when counting false positives in binary classification:

if y_pred.get(i) == y_true.get(i) {
    if classes == 2 {
        if *y_true.get(i) == T::one() {
            tp += 1;
        }
    } else {
        tp += 1;
    }
} else if classes == 2 {
    if *y_true.get(i) == T::one() {
        fp += 1;
    }
}

This checks if y_true == 1 when predictions don't match. However, this is counting false negatives (when true=1 but pred=0), not false positives. False positives should be when we predict 1 but true label is 0.

Similarly, in the Recall implementation, when counting false negatives:

if y_pred.get(i) == y_true.get(i) {
    if classes == 2 {
        if *y_true.get(i) == T::one() {
            tp += 1;
        }
    } else {
        tp += 1;
    }
} else if classes == 2 {
    if *y_true.get(i) != T::one() {
        fne += 1;
    }
}

This checks if y_true == 0 when predictions don't match, which actually counts false positives (when true=0 but pred=1), not false negatives.

Impact:

• Precision scores are actually Recall scores
• Recall scores are actually Precision scores
• F-beta scores with beta ≠ 1 are incorrect

[Mec-iS]

thanks for pointing this out.

it would be great if we can have a python script that demonstrates the discrepancy from scikit-learn using the same input.

If it is confirmed the fix should be quite straightforward, so anybody can open a PR with the fix