Open benwtrent opened 5 years ago
Pinging @elastic/ml-core (:ml)
A threshold would also give a similar user experience to outliers i.e. feature_influence_threshold
and potentially more useful results, although without a hard limit.
num_top_classes
allows a hard limit, but these could be pointlessly low probabilities.
Also to note, having two ways to set this does feel like overkill.
It is currently only possible to provide a static number for the number of top classes to return (
num_top_classes
).It would be beneficial to have a
top_class_threshold
or something similar so that all the top classes that are above a given threshold are returned.This feature would give some good indication (at a glance) around the distribution of probabilities for a given prediction. Additionally, if
top_classes
is empty on return, it could indicate a very low probability/confidence of the given predicted value.