Candidate Mapping at IV08
I’m here in London for the IV08 Information Visualisation conference where I’m presenting a paper titled “Candidate Mapping: Finding Your Place Amongst the Candidates.” This is the product of an independent study I did with Katy Börner. You can grab my author’s copy here, and can see my presentation slides here.
The process I followed essentially imposes a metric space on the non-metric information through the use of Gower Similarity equivalence classes. Unfortunately, the Gower similarity space is still incomplete. This is due to a few candidates neglecting to hold an opinion on some of the issues. This is handled through the use of Sammon Mapping. Sammon mapping seeks to reduce an error function based on the similarity of the entities.
What ends up happening is that candidates are “dumped” into a low dimensional space, and then “moved around” until they preserve the similarities they share with other candidates as best as possible.
For the most part, only a few of the candidates were missing many responsesn, such as Mike Gravel. In total, only 16 out of the 250 issue stances (25 candidates @ 10 issues each) were missing.
If the participant indicates his position according to the layout of map (which is generated specially for each participant and already includes his “actual position”), then the difference in position reflects a difference in similarity between the two positions, and not between every other candidate (as a dissimilarity). The distance between positions (selected by the user, and calculated by Sammon mapping) is not a true metric distance. However, since very few of the candidates are missing data, the Sammon mapping representation is nearly identical to a linear method using only candidates and issue stances with complete information. Therefore, distances as error are still informative. Other approaches for calculating error might involve calculating differences in stress (I’m continuing to look into this).
Finally, as a closing point, the final plot showing all the participants on the image at once is a composite of many different maps. Each of these maps will have slightly different layouts depending on how the participant answered (i.e. the position of the candidates will shift around slightly). However, the shifting is not enough to significantly change the general distribution of candidates, and so I chose a single configuration for the candidates, and indicated their general position using a large font.
Finally, the one issue that did not account for with this technique is that some people may have a “centrality bias” with regards to interaction with interfaces, as well as with reporting the overall “centrality” of their opinions. In other words, people may be choosing central positions just to avoid being on the “fringe” for one reason or the other, even though that is in fact their calculated position. I think this is very interesting because in spatial (directional) voting theory (which I reference often) includes a “region of acceptability”( check here for a recent paper explaining some of the concepts) parameter that accommodates this same centrality bias in voters. I.E., even though they endorse a set of ideologies that identifies them as “fringe,” they will not endorse comparable “fringe” candidates. This is an interesting social phenomenon that I’ll have to continue to look out for.
I also had several people come up to me and ask about the paper (also, someone wanted to know more about my groovy laTex-Beamer setup, find more about that here, and then you can check out the source for my presentation (without pictures/plots) here.
Thanks to Joseph Cottam for discussions.