I am currently working my way through the XGBoost algorithm and I had a question about the way in which the histogram is computed.
Each time we arrive at a new node we need to re-compute the histogram. However, I am not sure about one thing. Are the bins of the histogram computed only once at the top of the algorithm and then used every time that we create a histogram? Or is it the case that we re-compute the bins for the new histogram that we create? If the latter I don’t think I fully understand their advantage (unless we use some approximate method to compute the bins other than sorting the data, is this what’s happening? ).