General plan outline

[miguelgfierro]

Contents of MVP2:

• [x] Optimize airline experiment
• [x] Add comparison with XGBoost hist
• [x] Add Guolin experiments and run them on the same machine
• [x] Update notebooks with new parameters
• [x] Add experiments with GPU
• [x] Change notebooks with bokeh
• [x] Blog writeup

[miguelgfierro]

TODO cleanup:

• [x] airline: remove evaluation
• [x] football remove evaluation
• [x] remove xgboost hist and leave the loader for kaggle
• [x] change df_plane_numeric_small = df_plane_numeric.sample(n=10000000).reset_index(drop=True) to df_plane_numeric_small = df_plane_numeric.sample(n=1e6).reset_index(drop=True) in GPU airline

-remember gotchas: 1) evaluation will slow the training 2) custom loss functions in python are also slower 3) ligthgbm is cache aware

[miguelgfierro]

TODO GPU:

• finish parameters in both airline and higgs
• code results for different dataset sizes
• add GPU times and perf to the LightgbmvsXgboost notebook
• also the results for different dataset sizes

TODO bokeh:

• save as png and render it to github
• change requirements with recent version of bokeh

TODO blog:

• export bokeh plots to htlm+js
• try to render bokeh in the post blog
• think about the normal post
• think about Q&A
• decide if put Q&A together or in a different post