On Wed, Jul 14, 2010 at 01:21:19PM +0200, Yeb Havinga wrote:
> Heikki Linnakangas wrote:
>> However, the problem is how to represent and store the
>> cross-correlation. For fields with low cardinality, like "gender" and
>> boolean "breast-cancer-or-not" you can count the prevalence of all the
>> different combinations, but that doesn't scale. Another often cited
>> example is zip code + street address. There's clearly a strong
>> correlation between them, but how do you represent that?
>>
>> For scalar values we currently store a histogram. I suppose we could
>> create a 2D histogram for two columns, but that doesn't actually help
>> with the zip code + street address problem.
> In my head the neuron for 'principle component analysis' went on while
> reading this. Back in college it was used to prepare input data before
> feeding it into a neural network. Maybe ideas from PCA could be helpful?
I've been playing off and on with an idea along these lines, which builds an
empirical copula[1] to represent correlations between columns where there
exists a multi-column index containing those columns. This copula gets stored
in pg_statistic. There are plenty of unresolved questions (and a crash I
introduced and haven't had time to track down), but the code I've been working
on is here[2] in the multicolstat branch. Most of the changes are in
analyze.c; no user-visible changes have been introduced. For that matter,
there aren't any changes yet actually to use the values once calculated (more
unresolved questions get in the way there), but it's a start.
[1] http://en.wikipedia.org/wiki/Copula_(statistics)
[2] http://git.postgresql.org/gitweb?p=users/eggyknap/postgres.git
--
Joshua Tolley / eggyknap
End Point Corporation
http://www.endpoint.com