Thanks for sharing this. Of course I will closely watch it because claiming to beat gbdts might be a bit early.
- It is not entirely clear how the datasets split is done. Do you make sure that the model is evaluated on unseen data ? More generally how does one knows whether a dataset was part of the training or not ?
- You mention some serious limitations (10k rows, 500 cols.). It seems a bit weird to have fixed numbers. Can these numbers be roughly balanced ? (eg. 1M rows, 5 columns ... ). Does these numbers scale with memory ? (what memory was used for the 10k rows / 500 cols figure ?)
- It is not entirely clear how the datasets split is done. Do you make sure that the model is evaluated on unseen data ? More generally how does one knows whether a dataset was part of the training or not ?
- You mention some serious limitations (10k rows, 500 cols.). It seems a bit weird to have fixed numbers. Can these numbers be roughly balanced ? (eg. 1M rows, 5 columns ... ). Does these numbers scale with memory ? (what memory was used for the 10k rows / 500 cols figure ?)