Accessing the pol.is vizualization

[chrvt]

One can download the polis opinion matrix on which the clustering, dimensionality reduction, and low-dimensional visualization is based on.

Could I go the other way around? Given such an opinion matrix, could I use the pol.is algorithm to produce the low-dimensional visualization?

If so, what would be the steps to do so?

[metasoarous]

Hi @chrvt; Thanks for asking about this.

Yes, you can. You can take the matrix and perform PCA to obtain the two dimensional projection. There are some python notebook examples of this here:

• here: https://github.com/compdemocracy/notebooks
• and here; https://github.com/compdemocracy/analysis/tree/master/notebooks/jupyter (the above will likely get moved here)

that having been said, the exports themselves actually include the PCA projections, so you don't necessarily have to do this, unless you're looking to tinker.

[raulsperoni]

Hi! This is what I've been trying to do! But I'm struggling with this part:

"they're just projected according to the PC loadings, scaled by the same factor applied towards participant projections, so that they can be more easily plotted together"

I'm using sklearn PCA, fitting on the participants answers, and applying the fitted model to the "made up" points consisting in one answer per comment.

[patcon]

I don't understand this mathy magic that y'all do, but heads up that @metasoarous led a wonderful office hour session walking through some of the key concepts. So it might be covered in my unofficial notes from that: https://hackmd.io/@patcon/conversa-calls/https%3A%2F%2Fhackmd.io%2F%40patcon%2Fr1KpFakV_

[patcon]

Oops. The direct link wasn't working correctly in markdown, but I fixed it

[raulsperoni]

thanks! that link is dope! I will try to understand more

[phoqui]

that having been said, the exports themselves actually include the PCA projections, so you don't necessarily have to do this, unless you're looking to tinker.

Thanks for pointing this out, @metasoarous! Just to check, are you referring to the 'group-id' variable in the 'participants-votes' file? I'm assuming this shows what opinion group each participant has been assigned to? And where group-id = 0, does this mean the participant is an outlier i.e. they have not been assigned to an opinion group?
If you mean that there's actually data in the exports showing more specifically where each participant is projected on the 2D opinion landscape, please let me know - I haven't been able to identify anything like that so far.

[colinmegill]

Would welcome a PR of this to the python notebooks so that everyone can read through the logic. That will avoid ambiguity about implemention in the future.

[chrvt]

I started a version to show how groups are created and comment chosen according to the original paper, however, not all of the methodology is clear to me from, see https://github.com/compdemocracy/analysis/pull/4

[patcon]

Nice! Fixed link: compdemocracy/analysis#4