Open fabian-s opened 6 years ago
@jeff-goldsmith : any preferences / any experience which depths work well and should be the default?
i'd vote roahd::MBD.default
of these options, although i haven't used it much myself.
i guess we'll trust users to compute quantile / median curves for smooth feval_reg
data?
Even cheaper and more intuitive alternative: half region depth
I think there’s a super-computationally efficient implementation of MBD, but I’d have to dig around for it. Or by “cheaper” do you mean “fewer dependencies”?
On Mar 14, 2018, at 10:54 AM, Fabian Scheipl notifications@github.com<mailto:notifications@github.com> wrote:
Even cheaper and more intuitive alternative: half region depthhttps://www.sciencedirect.com/science/article/pii/S0167947310004123
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/fabian-s/tidyfun/issues/19#issuecomment-373050311, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AH_3hxsSTzspOqLsKHv2cp0Lno7d_Wduks5teS9CgaJpZM4SVFLu.
yeah, I know, roahd
uses the trick you probably mean. I did mean cheaper computationally, but come to think about it this might not actually be true....
Serfling, R., & Wijesuriya, U. (2017). Depth-based nonparametric description of functional data, with emphasis on use of spatial depth. Computational Statistics & Data Analysis, 105, 24-45. seems like a good resource for this
rank
/order
and proper min
/max
methods?More Lit:
Center-Outward Distribution Functions, Quantiles, Ranks, and Signs: https://arxiv.org/abs/1806.01238
Nadja&Thomas seem to have used some of this stuff in https://arxiv.org/pdf/1906.03151.pdf, ask them?
depths for irregular funs/ functional fragments: https://www.tandfonline.com/doi/abs/10.1080/10618600.2022.2070171
"random" depth: solid theory + comp. efficient (?) https://www.ma.imperial.ac.uk/~hbattey/STS532.pdf https://arxiv.org/pdf/2206.13897.pdf
see also experiments in attic/dev-quantile.Rmd
... not feasible for
feval_irreg
... for
feval_reg
: use MBD? (seeroahd:::MBD.default
, e.g.), other depths as inrainbow:::fdepth
,ddalpha::depthf
?... for
fbase
: could use multivariate depths on coefs directly (but consider scaling of basis functions)