Why scaleDims=TRUE do z-score normalization for each cell instead of each feature

[RoseYuan]

Hi all,

I'm looking at the function addIterativeLSI() and trying to understand what scaleDims = TRUE does to the feature matrix. If I understand it correctly, it calculates the z-score normalization for each row (i.e. each cell), according to the code here. I don't understand why one would do this. Let's say we have a matrix x of 3 cells x 3 features: x <- t(matrix(c(1,2,3,1,3,3,10,11,12),nrow=3)) wouldn't it be obvious that cell 1 and cell 2 are more likely to be from the same cluster than cell 1 and cell 3? But after applying the .rowZscores function, the matrix becomes like this,

-0.5773503 | -0.5773503 | 1.154701
-0.6757374 | -0.4730162 | 1.148754
-0.5773503 | -0.5773503 | 1.154701

suggesting that cell 1 and cell 3 are from the same cluster... So I'm confused why this normalization could be good.

[dannyconrad]

I just raised this concern as well at #2120 before I saw this thread. I agree this probably needs to be addressed.