@mrohban I removed the random projections part in covariance profiling #114.
I wondering whether we should have random projections as a separate feature transformation step? That's because if you have say, just 10 variables, it may be ok to have 45 covariance features.
The disadvantage of that approach is that for larger feature sets, the dataframe can get too wide, take up memory, making it a pain to do random projections later.
@mrohban I removed the random projections part in covariance profiling #114.
I wondering whether we should have random projections as a separate feature transformation step? That's because if you have say, just 10 variables, it may be ok to have 45 covariance features.
The disadvantage of that approach is that for larger feature sets, the dataframe can get too wide, take up memory, making it a pain to do random projections later.
One mitigant for this is to do sparse random projections, which will be faster.
What do you think?