Plot recipe for MDS from MultivariateStats.jl

[kescobo]

This is dependent on JuliaStats/MultivariateStats.jl#85, though I'm not quite sure how to indicate that in compat until it's tagged.

using StatsPlots
using Distances
using MultivariateStats

m = rand(100, 10)

dm = pairwise(BrayCurtis(), m)

mds = fit(MDS, dm, distances=true)

plot(mds)

[mkborregaard]

That looks sweet! Have you considered an interface for multiple dimensions? Either passing dimensions manually, or using subplots or possibly 3d?

[kescobo]

@mkborregaard I have considered, but wasn't sure about the implementation. Since this is not a user recipe, can I add a keyword for that?

[mkborregaard]

Sure :-)

[mkborregaard]

(technically this is a user recipe)

[kescobo]

I think I was mis-remembering the deal with keyword arguments...

In any case, this now allows a user to specify axes, and do a 3rd axis:

using StatsPlots
using Distances
using MultivariateStats

m = rand(100, 10)
dm = pairwise(BrayCurtis(), m)
mds = fit(MDS, dm, distances=true)

plot(mds)

plot(mds, mds_axes=(3,5))

plot(mds, mds_axes=(1,2,3))

[mkborregaard]

Love it ❤️

[mkborregaard]

Oh - I merged already, but only now realised - shouldn't aspect_ratio have been set to 1?

[mkborregaard]

Also, I think you're missing a dep on MultivariateStats for this to work?

[mkborregaard]

Wait, you do have that, I seem to have a local problem, sorry for the noise

[kescobo]

Not sure, what does that do? Happy to add it in.

[mkborregaard]

@kescobo I'm getting WARNING: could not import MultivariateStats.MDS into StatsPlots locally - is there something wrong with my setup/versions?

Also, could the signature be wider to allow for PCAs and other ordinations as well? Is the interface not general enough for this?

It would be awesome to have a readme example with this code

using MultivariateStats, RDatasets, StatsPlots

iris = dataset("datasets", "iris")
Xtr = convert(Matrix, iris[:, 1:4])'
M = fit(PCA, Xtr; maxoutdim=2)
plot(M, group = iris.Species)

or something like that. Compare the code here: https://multivariatestatsjl.readthedocs.io/en/latest/pca.html#data-analysis

[mkborregaard]

Not sure, what does that do?

Forces the axes to have the same scale. Given that they are by definition on the same scale it gives a more correct representation of the spread on the axes.

[kescobo]

@kescobo I'm getting WARNING: could not import MultivariateStats.MDS into StatsPlots locally - is there something wrong with my setup/versions?

You have to be on master of MultivariatStats (see the first comment in this thread=). The PR introducing this API hasn't been tagged, so I can't add it to [compat] in the Project.toml yet (I asked on slack)

Also, could the signature be wider to allow for PCAs and other ordinations as well? Is the interface not general enough for this?

I haven't looked into it, but I thought about the same thing. I will happily do a separate PR.

It would be awesome to have a readme example with this code

Sure - should this be a separate PR since this one is already merged?

[kescobo]

Forces the axes to have the same scale. Given that they are by definition on the same scale. Given that they are by definition on the same scale it gives a more correct representation of the spread on the axes.

My understanding was that for PCoA at least, the scales are essentially unitless and so scaling is meaningless. Entirely possible I misunderstood that though 😆

[mkborregaard]

You have to be on master of MultivariatStats (see the first comment in this thread=)

Sorry 😂

I will happily do a separate PR.

Yay! 💖

should this be a separate PR

I think it has to be yeah.

scales are essentially unitless

AFAIK the PCoA transformation maintains distances among points. But if Plots chooses different scales for the x and y axis (based on the range of the data and the shape of the plotting window), the distances will not be correct.

[kescobo]

AFAIK the PCoA transformation maintains distances among points.

But PCoA is for non-metric distances right? Maybe I'll ask on Slack...