Do you have any idea what the root of this could be?
Maybe the error is on our side?
You can find our implementation in Filter.R.
PS: It is very nice that praznik is so fast!
AFAICT, mRMRe calculates mutual information by estimating from correlation between features, using the formula I(x,y)=−.5 log (1−cor(x,y)^2) (this is why it expects numerical input), while praznik from a direct formula I(x,y)=p_xy log (p_xy/p_x/p_y) for categorical variables (and this is why it expects factors)... So these are basically different algorithms, or let's say, versions of mRMR using different interface to the data.
I(x,y)=−.5 log (1−cor(x,y)^2)
I(x,y)=p_xy log (p_xy/p_x/p_y)