I did not like the comment. It is needlessly aggressive and implies I have the arrogant view that my post is better than the author's paper. I do not.
I also suspect your hypothesis is not correct. The top row of your link is a good example. The NN tries to infer a gradient but that's really tough to do with limited data. That is to say, the tree based models will locally fit their partitions to the exact training data and the NN will try to view it in the big picture filling in the gaps. Tree based model works better for most real world tabular data.
The paper concludes something similar:
> This superiority is explained by specific features of tabular data: irregular patterns in the target function, uninformative features, and non rotationally-invariant data where linear combinations of features misrepresent the information.
I daresay my perspective is better aligned with the paper than yours. Are YOU trying to replace the publication with your 4 lines?
Totally agree and sorry again for my wording as I probably miss understood what you meant.
And no, I am trying to replace what I consider is a wrong intuition (tree methods are strong models mostly because single trees separate data in hyperplanes) with my 4 lines. This is just my opinion.
I also suspect your hypothesis is not correct. The top row of your link is a good example. The NN tries to infer a gradient but that's really tough to do with limited data. That is to say, the tree based models will locally fit their partitions to the exact training data and the NN will try to view it in the big picture filling in the gaps. Tree based model works better for most real world tabular data.
The paper concludes something similar:
> This superiority is explained by specific features of tabular data: irregular patterns in the target function, uninformative features, and non rotationally-invariant data where linear combinations of features misrepresent the information.
I daresay my perspective is better aligned with the paper than yours. Are YOU trying to replace the publication with your 4 lines?