GBDT demo for sample train
Here is the reference blog: http://nbviewer.jupyter.org/github/liudragonfly/GBDT/blob/master/GBDT.ipynb
innitialize F for each epoch: 1. cur_epoch_samples= 0.8* total_samples 2. compute residual of every sample (equation 14) 3. build tree based on MSE, and then compute the weight of each node (equation 15-2) 4. update F value with weights of sample-nodes and other samples