You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
As I recall, we allow the specification of a target y when training an unsupervised detector, even though it is ignored, so that we may buy into MLJ's evaluate apparatus: If you have test labels, but no train labels, you just pad y with missings in the training indices.
It seems to me that the target_scitype should therefore be AbstractVector{<:Union{OrderedFactor{2},Missing}} in those cases, but I see that it is actually AbstractVector{OrderedFactor{2}} (at least it is for ABODDetector that I have been playing with). The bottom line is that MLJ is complaining when checking scitypes for these models, if the target presented has missings. (Actually, MLJBase <= 0.19.8 is just throwing an error, because of a bug; MLJBase >=0.20 would throw a warning, I guess, if it was compatible with OutlierDetection, which is isn't currently.)
@davnn Do you see any reason not to expand the scitype?
Hi, I recall it as a temporary fix, because missing values were causing some problems when using other MLJ features, e.g. ensemble models, but I think you have addressed that in JuliaAI/MLJEnsembles.jl@3e55cb6. Otherwise, I think we can expand the scitype and fix potential problems in MLJ down the line.
As I recall, we allow the specification of a target
y
when training an unsupervised detector, even though it is ignored, so that we may buy into MLJ'sevaluate
apparatus: If you have test labels, but no train labels, you just pady
withmissing
s in the training indices.It seems to me that the
target_scitype
should therefore beAbstractVector{<:Union{OrderedFactor{2},Missing}}
in those cases, but I see that it is actuallyAbstractVector{OrderedFactor{2}}
(at least it is forABODDetector
that I have been playing with). The bottom line is that MLJ is complaining when checking scitypes for these models, if the target presented hasmissing
s. (Actually, MLJBase <= 0.19.8 is just throwing an error, because of a bug; MLJBase >=0.20 would throw a warning, I guess, if it was compatible with OutlierDetection, which is isn't currently.)@davnn Do you see any reason not to expand the scitype?
@josephsdavid
The text was updated successfully, but these errors were encountered: