Make num_top_classes parameter's default value equal to 2#48119
Make num_top_classes parameter's default value equal to 2#48119przemekwitek merged 3 commits intoelastic:masterfrom
Conversation
|
Pinging @elastic/ml-core (:ml) |
549320b to
5d6c1af
Compare
2e5121b to
98240f7
Compare
There was a problem hiding this comment.
This does not really make sense for multi-class as returning the top two will seem arbitrary in that situation.
Maybe just the probability of the one we return (i.e. numTopClasses = 1)?
One could see a parallel between this setting and the computeFeatureInfluence and featureInfluenceThreshold settings for outlier_detection.
It almost seems like we need a "close to" parameter that returns all the top classes that are above a certain probability... @tveasey what say you?
I agree. Please remember though that in 7.5 we don't go multiclass and for binary "2" is a sensible default.
Let's revisit it in 7.6 when we go multiclass.
|
benwtrent
left a comment
There was a problem hiding this comment.
I am OK with this, we just need to document that the default value is 2 somewhere and that supplying a 0 is needed to prevent them from being returned.
I do think that a parameter like top_classes_threshold for the user to provide a probability threshold and return the top classes that are above that threshold would be a great addition. I will open an separate issue for that :)
|
run elasticsearch-ci/packaging-sample |
1 similar comment
|
run elasticsearch-ci/packaging-sample |
422ab37 to
dbefbc9
Compare
Document this parameter in HLRC docs
dbefbc9 to
0c04ba7
Compare
|
run elasticsearch-ci/packaging-sample-matrix |
Users are often interested in the probabilities of each class being predicted.
This PR sets the default of
num_top_classesparameter (which controls how many classes with probabilities are returned) to 2.Relates to #46735