Currently
julia> MeanStdScaling([1])
MeanStdScaling(1.0, NaN)
due to Statistics.std behaviour: corrected=true by default, which divides the sum of errors by N-1 instead of N, and N-1 is 0 for singletons.
It's an edge case, but could arise in feature engineering pipelines where the size of available data varies. I think for the purpose of feature engineering, a singleton vector should be standardized to [0] rather than [NaN].
Currently
due to
Statistics.stdbehaviour:corrected=trueby default, which divides the sum of errors by N-1 instead of N, and N-1 is 0 for singletons.It's an edge case, but could arise in feature engineering pipelines where the size of available data varies. I think for the purpose of feature engineering, a singleton vector should be standardized to
[0]rather than[NaN].