ENH: add normalize parameter to metrics.classification.confusion_matrix #14478
Conversation
Allows to get a normalized confusion matrix directly from the function call. I use this function frequently and find the need to always normalize the matrix manually a bit tedious.
| @@ -211,6 +212,10 @@ def confusion_matrix(y_true, y_pred, labels=None, sample_weight=None): | |||
| If none is given, those that appear at least once | |||
| in ``y_true`` or ``y_pred`` are used in sorted order. | |||
|
|
|||
| normalize : bool, optional (default=False) | |||
| If True, , return the fraction of classified samples (float), | |||
There was a problem hiding this comment.
You can normalise by number of samples (like accuracy), by number of samples in the ground truth for each class (like recall), or by number of samples predicted for each class (like precision). I find this description highly ambiguous, and I'm not persuaded we should assume one normalisation is more appropriate than another for the user.
There was a problem hiding this comment.
Good point. I did it that way because most frequently the normalization is done with regard to the ground truth. When we think of a normalized confusion matrix the picture that comes to mind is precisely something like this.
To avoid suggesting to the user just one specific type of normalization, different arguments for the normalize parameter could be specified, e.g, "precision/accuracy/recall".
|
I'm okay with giving options. I think calling it precision/recall/accuracy
is a bit misleading since they don't pertain off the diagonal of the
matrix. true vs pred might be better names. It's still not entirely clear
to me that providing this facility is of great benefit to users.
|
| @@ -184,7 +184,8 @@ def accuracy_score(y_true, y_pred, normalize=True, sample_weight=None): | |||
| return _weighted_sum(score, sample_weight, normalize) | |||
|
|
|||
|
|
|||
| def confusion_matrix(y_true, y_pred, labels=None, sample_weight=None): | |||
| def confusion_matrix(y_true, y_pred, labels=None, | |||
| normalize=False, sample_weight=None): | |||
There was a problem hiding this comment.
You should keep sample_weight before normalize just in case.
|
With your proposal, you also need to implement tests to ensure that the function will work properly. |
Allows to get a normalized confusion matrix directly from the function
call. I use
confusion_matrixfrequently and find the need to alwaysnormalize the matrix manually maybe unnecessary.
I am aware of the fact that other functions like
accuracy_scorealreadyhave this exact functionality implemented, so probably the lack of the
normalizeparameter is intentional and I'm missing the why. But in caseits not intentional you might find this contribution useful :).