I am computing precision and recall for my classification and I am finding that the recall value for some classes is higher than the precision value. I am curious if that's even possible.
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
It is totally possible, they are different things.
Here is an example: suppose we are doing fraud detection.
- There are 10 fraud cases in 1000, data points.
- The model is a dummy model that report all 1000 cases are fraud.
In such example, the recall is 100% and precision is 0.1%
answered Oct 12, 2017 at 16:50
-
$\begingroup$ What can one do to turn this around (especially for recommendation problems)? My assumption is that positive sample set is not good and negative sample set is better and more generalized. $\endgroup$– SadeCommented Mar 3, 2021 at 8:51
-
$\begingroup$ For an image classification model where target is cat or not cat (for example), it much easier to resolve by changing the threshold and you can directly affect the precision score. $\endgroup$– SadeCommented Mar 3, 2021 at 8:55
-
$\begingroup$ I think that @Haitao Du's answer makes sense, except that the precision should be 1% instead of 0.1%. $\endgroup$ Commented Mar 11, 2022 at 12:07