#new sota for tabular data ?
bright pumice
A team from yandex (russian search engine) has published a paper about a new kind of model. The idea as I understand it is to integrate some form of knn in the architecture. They claim to beat state of the art gbdt on tabular data. Claim is pretty serious as 1) they already author the catboost package and 2) the evaluation is done on a serious benchmark (the one used in the paper about why gbdt still outperform nn). Feel free to give some feedback especially if you managed to try it.
new sota for tabular data ?
quick perch
Xgboost is the only model you could ever possibly need - bojan
scenic thunder
It is a very promising model from the team who are experts in DL for tabular datasets.
Definitely give it a try. Unfortunately from a practical point of view it is prohibitively slow for datasets more than a few 100K samples. They admitted this in the paper.
dense crag
Maybe there's a tabular playground series dataset it could work well to test on
bright pumice
Have this new technique been tested ?