You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@commons.apache.org by Phil Steitz <ph...@steitz.com> on 2009/02/08 21:52:35 UTC
[math] Correlation and Covariance
MATH-114 and MATH-138 propose support for correlation matrices. I have
been working on these and would like to propose the following:
Create a new package o.a.c.m.stat.correlation to house intially
a) Covariance - creates variance-covariance matrix from a matrix
whose columns represent covariates. Also includes convenience methods
that work pairwise on double[] arrays (similar to VectorialCovariance,
but requiring that the arrays be stored)
b) PearsonCorrelation - creates Pearson's product-moment correlation
matrix from either a covariance matrix or a matrix of covariates. Also
includes methods to return matrices of correlation standard errors and
p-values (aka significances, i.e. p-value for null hypothesis that the
coefficient is 0).
c) SpearmanRankCorrelation - like Pearson's but no covariance matrix
constructor and using rank correlation.
To implement c), we need a place for the RankingAlgorithm interface and
implementations (see MATH-138). Any suggestions on where to put
these? Leaving in correlation may be awkward later on as we do more
with rank transformations.
I have a) implemented using a fairly stable two-pass algorithm. I tried
just using VectorialCovariance, but could not get the accuracy I wanted
using the one-pass algorithm there. We should probably at some point
look at improving the updating formula used there along the lines of
what we do for Variance, but it is a nice feature of that class that it
does not require the input vectors to be stored and I would not want to
see that changed. For b), similar to the patch in JIRA, I would use
the R computation from SimpleRegression if working from a matrix, or
just compute column sigmas and scale directly if working from a
covariance matrix.
Does this sound good?
If I don't hear any objections, I will commit some code along the lines
above for us to look at.
Phil
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org
Re: [math] Correlation and Covariance
Posted by Luc Maisonobe <Lu...@free.fr>.
Phil Steitz a écrit :
> MATH-114 and MATH-138 propose support for correlation matrices. I have
> been working on these and would like to propose the following:
>
> Create a new package o.a.c.m.stat.correlation to house intially
> a) Covariance - creates variance-covariance matrix from a matrix
> whose columns represent covariates. Also includes convenience methods
> that work pairwise on double[] arrays (similar to VectorialCovariance,
> but requiring that the arrays be stored)
> b) PearsonCorrelation - creates Pearson's product-moment correlation
> matrix from either a covariance matrix or a matrix of covariates. Also
> includes methods to return matrices of correlation standard errors and
> p-values (aka significances, i.e. p-value for null hypothesis that the
> coefficient is 0).
> c) SpearmanRankCorrelation - like Pearson's but no covariance matrix
> constructor and using rank correlation.
> To implement c), we need a place for the RankingAlgorithm interface and
> implementations (see MATH-138). Any suggestions on where to put
> these? Leaving in correlation may be awkward later on as we do more
> with rank transformations.
>
> I have a) implemented using a fairly stable two-pass algorithm. I tried
> just using VectorialCovariance, but could not get the accuracy I wanted
> using the one-pass algorithm there. We should probably at some point
> look at improving the updating formula used there along the lines of
> what we do for Variance, but it is a nice feature of that class that it
> does not require the input vectors to be stored and I would not want to
> see that changed. For b), similar to the patch in JIRA, I would use
> the R computation from SimpleRegression if working from a matrix, or
> just compute column sigmas and scale directly if working from a
> covariance matrix.
This seems good to me. Perhaps a dedicated "ranking" package under stat
would be fine.
Luc
>
> Does this sound good?
>
> If I don't hear any objections, I will commit some code along the lines
> above for us to look at.
>
> Phil
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
> For additional commands, e-mail: dev-help@commons.apache.org
>
>
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org