Don't forget that "time of year when photo was taken" affects the results -- snowy scenes taken might beat "photo of a road on a dull drizzly day, taken only to fill up the Geograph grid", and whilst there is guidance to not rate the photographer, a good composition will enhance the view;
As an example; two photographs taken at the same locale, by the same person, within a few minutes of each other, pointing in different directions: http://www.flickr.com/photos/ratarsed/3377071312/ http://www.flickr.com/photos/ratarsed/3376131577/ Both photos have their artistic merits, but in my mind, one is a lot more scenic than the other -- so how should this affect the ranking of that area? (For reference, if I saw these on ScenicOrNot, I'd rate one as a 3, and the other as an 8) 2009/4/8 Francis Davey <[email protected]>: > 2009/4/8 Frankie Roberto <[email protected]>: >> >> I'd be fascinated to know how a factor analysis works (I tried looking at >> http://en.wikipedia.org/wiki/Factor_analysis, but it's not the most >> accessible Wikipedia page). > > No, its awful. > > I'm using the term a bit generically but its quite simple. > > Eg, imagine there are N people who have voted on pictures. Now take an > N dimensional graph and plot where they rate them all (or how they > compare them all). Each picture is a point in this N-dimensional > space. > > Now we have an utterly incomprehensible graph which is also hard to > visualise to those of us who find thinking in more dimensions than we > have toes difficult. > > So, what would be great is to somehow reduce that number of dimensions > a bit, or even a lot. That amounts to finding a few factors that > explain most of the data. > > How you do this, like much of stats, depends. There are lots and lots > of algorithms for it. Some are easy - roughly corresponding to > projecting the N-dimensional space down onto some subspace that's more > manageable, so all you have to do is find the subspace. But there's no > reason to assume that everything is linear, so you might do something > more sophisticated. > >> >> Another alternative might be to force people to make a binary choice between >> "scenic" and "not scenic", or perhaps a 4 way choice with 2 "very" options. >> Then you avoid all the indecisive 4-6 responses. >> > > If what you want is a *lot* of data comparisons fast then use > something like Maxdiff: > > http://en.wikipedia.org/wiki/MaxDiff > > Show four photos and ask for best and worst. That's still amazingly > easy (almost as easy as the kittenwar game) and you get a lot more > ratings done. > > But, beware! In this case there's a whole nother issue. So far we have > been considering: > > - finding scenic places on the basis of some mass voting (a million > people can't be wrong) > - finding places I'd like (needs a factor analysis or something similar) > > But the scenes have location data too. You might want to say here -> > is an really good place to go because there is a cluster of scenically > rated photos from there. That requires a whole lot more sophisticated > analysis again. > > However I don't know what the use cases of this data might be, so > can't comment. I'm not saying Tom et al. are wrong because they know > what their constraints and aims are which I most emphatically do not. > What's more they have almost certainly taken the advice of > statisticians to get this just right, so my rather amateurish > criticism is meant to be just that, my half pennyworth. > > When I get stuck, I tend to go off and talk to a fellow of the royal > statistical society. It tends to unstick my mind, though I usually > come away realising how much more problematic everything really is > 8-). > > -- > Francis Davey > > _______________________________________________ > Mailing list [email protected] > Archive, settings, or unsubscribe: > https://secure.mysociety.org/admin/lists/mailman/listinfo/developers-public > _______________________________________________ Mailing list [email protected] Archive, settings, or unsubscribe: https://secure.mysociety.org/admin/lists/mailman/listinfo/developers-public
