Found 2 comments on HN
jmalicki · 2014-12-19 · Original thread
RandomForest and other decision tree methods can actually handle missing data very well by treating an individual cell in the data matrix as missing, rather than discarding the entire row.

So the assertion that "All algorithms should operate only on data vectors and on frequency weights — they should have no knowledge of missing-ness." is false - there are a lot of other fruitful ways to handle missing data, such as using indicator variables, imputation, etc. - see

Get dozens of book recommendations delivered straight to your inbox every Thursday.