This is long overdue, but over the winter break we were finally able to write up our sgdGMF paper: arxiv.org/abs/2412.20509 We present a stochastic gradient descent method that allows to efficiently and very quickly estimate latent factors for, e.g., dimensionality reduction of single-cell data - ThreadSky

davi1893.bsky.social • 37 days ago

This is long overdue, but over the winter break we were finally able to write up our sgdGMF paper:
https://arxiv.org/abs/2412.20509

We present a stochastic gradient descent method that allows to efficiently and very quickly estimate latent factors for, e.g., dimensionality reduction of single-cell data

Comments

goringennady.bsky.social•36 days ago

Could you speak in a bit more detail to how the matrix W is determined? Near Eq. 2, it seems to be described as a user-set parameter, but it seems quite nontrivial to specify it without solving a much more challenging problem. But constraining it in an arbitrary way seems overly restrictive.

davi1893.bsky.social•36 days ago

Sure. If I understand your question, w_{ij} in eq.2 are optional known weights (as in weighted regression) that can be used if appropriate in the application or can be safely ignored (that's what we do in our single-cell examples in the paper). This isn't different from the basic lm() function.

goringennady.bsky.social•36 days ago

Exactly, those. It seems really restrictive to throw away or treat as arbitrary dispersion differences between cells (regrettably common) and between genes (less common, and oversimplified on its face, I think)

davi1893.bsky.social•36 days ago

Right. We currently have a global dispersion parameter. In our comparisons, this is enough to correctly estimate the latent factors (see also https://www.nature.com/articles/s41467-017-02554-5).

goringennady.bsky.social•36 days ago

Sure. My reservation is just that baking pretty severe and unusual misspecification into the model might make those unreliable or unstable. The same goes for ZI models (which, as far as I know, still have no mechanism or direct evidence).

davi1893.bsky.social•35 days ago

I fail to see how this is unusual, it's a standard model. As for reliability and stability, I believe that we make a good case for them in the paper for our model. I have my own thoughts on ZI models but I'll save them for another time as here we do not model ZI.

davi1893.bsky.social•36 days ago

Of course it could be useful in other applications (e.g., DE) to have genewise dispersions and it's something that we could add in a future version of the package.

davi1893.bsky.social•37 days ago

This work was driven by Cristian Castiglione and was a great collaboration with Lieven Clement and Alex Segers

ppaxisa.bsky.social•37 days ago

Congrats! Can’t wait to test it!

Comments

Posting Rules

Reply