I like the name bu, but I called this User Stylometry Association, or UStylA, in my paper. In short, this just clusters users based on their stylometry - how they write stuff. This ended up as my Senior Honours project at The University of St Andrews. I had more ambitious plans but I didn't have enough time for them. This isn't half bad either though.
Okay, so I was right. In my case the index of each of the probabilities doesn't relate to the actual class. So the first index could relate to the class 6 (you know, each class is a UserID) instead of 1. I'm definitely going to need a mapping to sort this out, I could either map each class to their actual index, or I could just straight up map them to their probabilities. The only problem with the probability map is that multiple classes could have the same probability (wait this might be fine and not a problem), but the index map might not work since the prediction array doesn't seem to use all of the classes...