Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!news.kei.com!world!mv!barney.gvi.net!redstone.interpath.net!sas!mozart.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Q: Stat. correlation and Input selection (time series forecasting)
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <DKonv3.27s@unx.sas.com>
Date: Fri, 5 Jan 1996 00:54:39 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References: <4capmb$69d@otis.netspace.net.au> <4cbnjo$f66@canton.charm.net>
Organization: SAS Institute Inc.
Lines: 34


In article <4cbnjo$f66@canton.charm.net>, Ken Otwell writes:
|> >   luke@ice.net.au (Luke) writes:
|> >  Can a net extract relationships from variables that appear to have
|> >  no correlation with the output?

Even linear regression can do that.

|> >  At the moment, I check the correlation coefficient between each input
|> >  and the output before including it in the model.   Could I be excluding
|> >  important variables with this approach, or is the existence of linear
|> >  correlation between the inputs/output a prerequisite?
|> >
|>
|> Well, if it were, we'd all still be doing linear regression. You might want
|> to try higher-order terms in your correlation filter (eg., add the
|> product of pairs of inputs as new inputs, then triples, etc., till you can't
|> stand it no more), and only exclude those inputs that show no correlation
|> in any combination. Of course, after all that work, you might just hand craft
|> a high-order model and stick with linear regression anyway...
|>
|> This doesn't detect all the nonlinearities that an NN might (such as nonmonotonic
|> effects), but its better than what you're doing now.

To check for nonlinear effects, you can discretize the inputs and
targets (if they're not categorical to begin with) into maybe 10
categories, crosstabulate them, and compute any of the usual statistics
for measuring association in crosstabulations.

-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
