> B) gather a full matrix of the level of "correlation" between each column and
> each other column. If this were a single floating point number per pair
> then it might be feasible. It would still obviously be n^2 in the number of
> columns though, so there would have to be some way to limit on how many
> columns would be analyzed this way.
Use foreign keys to just record those cross-correlations.
Chris
