The problem is that when blurring is done on a layer that includes transparent areas, it needs some information about the area surrounding the central object, but the algorithm has no idea what you might eventually put under it, so it has to make a guess.
It looks like it assumes the object is on a white background, blurs both the object and the white area, and then uses the original transparency info contained in the layer to feather the effect to transparency. If what's actually under the object is light, this is not a bad assumption. Otherwise, it's a bad assumption.
I think you are stuck applying the effect on composite layer of your own making, then applying the original transparency info to it. I know it's a manual, time intensive method, but I don't see any way around it.
T