At the very core of things, the computer only understands 0s and 1s haha. For human interpretability - yes, it might be necessary to label the columns to see what a specific missing indicator column is for, but for a computer, it doesn't matter. The column headers are just for us, as the model only cares about the data being in a 2D array format. For: A) the model doesn't know that a particular column is indicating missing values in some other column. It only cares about the values in it. B) Again, to reiterate, the model doesn't care what each of the other 93 odd columns stand for. It is only looking at their values You can shuffle the column ordering, pass in X.values instead of a dataframe X to the model. It will not affect performance

@@rachittoshniwal Yeah I understand that but the 93 columns are all features in and of themselves, the 4 imputation indicators aren't really features of data in the same underlying way, right? They are more like features about other features, not features about the event whose label it is trying to train on. It seems like the imputation indicator data point's entire utility is essentially to point at a single data point in another column and say "don't take this data point too seriously because it was made up" -- wouldn't the entire weighting system need to treat that types of columns be different? It feels like it would be problematic (or at least, not useful) to treat those columns as if they were just additional features columns that could be treated like any other of the existing 93 features. (I mean, I guess it depends on the particular algorithm, like i imagine decision tree based algos would probably be able to handle that kinda thing, but others it feels like wouldn't be well served to treat those columns like they were just any other feature columns, no different than any of the others for its own starting purposes)

@@subtlehyperbole4362 although the missing indicator column is based off some column X, it is essentially a brand new column holding the information that "there is a column X which has missing values for these rows" so the model will test whether NOT having values in that column X is indicative of something or not.

even if you apply scaling after encoding, the 0s and 1s will be converted to some new numbers, but all 0s will be the same number x and all 1s will be the same number y. So it is again essentially encoded, just that instead of 0s and 1s you have x and y

Thanks! I do have a couple of end to end project videos where I've fitted models after transforming. Hope they help!