Thoughts on adding SVM as a classifier?

[derrik-gratz]

I've enjoyed using the celltypist interface, but I've seen a few papers mentioning SVM outperforming other classification methods for cell annotation. I'm curious if there was motivation to exclude this from the package? I'm testing adding it myself as an extension of this package to utilize the nice models and interface of celltypist, but I thought I'd see if you had already considered this and opted against it.

https://academic.oup.com/bib/article/25/5/bbae392/7730135
https://genomebiology.biomedcentral.com/articles/10.1186/s13059-019-1795-z

[ChuanXu1]

@derrik-gratz, currently SVM classifiers are not included in CellTypist, mainly because SVMs do not provide native probability outputs. To obtain probabilities, an additional calibration step (e.g., fitting a logistic curve on top of the SVM decision scores) is required, whereas logistic regression produces probabilities naturally. However, from a pure classification perspective, SVMs can be a strong alternative to logistic regression. At the moment, we don’t have plans to add an SVM option to celltypist.train, but we may consider supporting it in the future.

[derrik-gratz]

Thanks for the response. I'm an ML novice, so there may be reasons that this method is not preferred, but SKLearn's SVC implementation has an option to return probability via cross validation.. I've tested this with cell annotation and it works as expected to return per-sample per-class probabilities that you can use to create thresholds for no-calls etc.