how tabnet with perform in the p >> n situation?

[parsifal9]

Hi Tabnet,

Do you have any insight on how tabnet with perform in the p >> n situation?

I have done the tutorial at https://blogs.rstudio.com/ai/posts/2021-02-11-tabnet/ with a data set of 11M by 29. However the use cases I am intersted in are in genomics where the data sets are closer to the transpose of that.

I have tried a small test data set is 150 x 14000 and in this case the runtime grows a lot. Do I have any hope? What about future developments?

Bye

[cregouby]

Hello @parsifal9,

I'm not from the single-cell genomic business, but I think tabnet cannot do that. This is because, by design, there is far too much degree of freedom in the tabnet network parameter to train on 150 samples.

[cregouby]

and to my knowledge, the Seurat methodology can not be challenged so easily ;-)

[cregouby]

but in case you have a very large number of observations n without outcome, you can try to pre-train the tabnet model with them. That maybe could help the convergence with a tiny supervised data budget...