foursquare analysis

[tmoon]

http://www.urbagram.net/archipelago/

via: http://gawker.com/5634150/your-citys-foursquare-fingerprint

links to data and analysis at bottom

[tmoon]

There was something that bugged me about this and now I kind of know
what it is. If you download the data and look at it, you'll see that
the coordinates are given to something like 10ish significant digits.
This is obviously necessary since it's concentrating on specific
areas, NYC, Paris and London. And so most of the coordinates fall into
a square region of maybe .2ish degrees.

This is going to be a problem for the spherical model I'm working on.
For data from all over the world, maybe 2 significant digits might be
enough to represent the world, but if we have a large set of points
concentrated around a small area, this is not going to be enough and
this presents problems in terms of (1) catastrophic cancellation if I
want increased precision since I'm subtracting as often as I'm adding
to some mean coordinate (2) kappa will have to be very big and I don't
know what the practical implications will be; most of the stuff I've
seen so far with spherical distributions seem to assume smallish kappa
(20, which is what I'm using right now, would be on the higher end).

After we're done experimenting with the first spherical model, I think
we'll have to build a model that does local, linear approximations
with a tangent plane (say, it's a vanilla gaussian distribution for
the north eastern seaboard).