| Re: regression optimization [message #71391] |
Thu, 17 June 2010 01:36 |
Klemen
Messages: 80
Registered: July 2009
|
Member |
|
|
On Jun 17, 10:01 am, Dave Poreh <d.po...@gmail.com> wrote:
> On Jun 17, 12:35 am, Klemen <klemen.zak...@gmail.com> wrote:
>
>
>
>> Hello all,
>
>> I have a variable to be explained with 12 attributes (at the moment
>> all of them being arrays of 300 by 100, but in the future it can be
>> even larger than 1000 by 1000 array). I would like to preform a moving
>> window analysis. For each window I want to estimate the optimal
>> multiple linear regression parameters.
>
>> Optimal means that it is not necessary that all of 12 variables are
>> used (some of them are correlated). So I want to get out an equation
>> that has 2,3,4... parameters and provide the best statistics.
>
>> The problem is also that not all the values within a moving window can
>> be used - some pixels contain no data. Defining NaN for these values
>> and using the code similar to the one written by Bringfried Stecklum
>> ( http://groups.google.com/group/comp.lang.idl-pvwave/browse_t hread/
>> thread/17613c70b78f1ac4/6891d260db6c7c93?
>> lnk=gst&q=regression#6891d260db6c7c93), I can test which of attributes
>> might be significant for the regression.
>
>> The question is, how to proceed. From those attributes that I know
>> that they are correlated among each other, I want to use just the one
>> that explains the most variability. I can somehow imagine to select
>> the final 2-4 attributes that should be used for multiple regression
>> without using any FOR loops. But how do I do the final step -
>> estimation of multiple regression parameters without using any loops?
>> Any idea?
>
>> Thank you!
>> Klemen
>
> I think the problem that you mentioned is a principle component
> analysis problem (PCI). For each state you have to apply this analysis
> and then select the first, second or what ever variables you need to
> justify the total variance. David has a very good example on it
> (http://www.dfanning.com/code_tips/pca.html).
> Cheers
> Dave
Hi Dave, thank you for your suggestion. I will think of it. I have
never used PCA before. I am just afraid that computing for each moving
window a 13 by 13 covariance matrix and its eigenvectors is also not
really to easy to write without any for loops.
Klemen
|
|
|
|
| Re: regression optimization [message #71394 is a reply to message #71391] |
Thu, 17 June 2010 01:01  |
d.poreh
Messages: 406
Registered: October 2007
|
Senior Member |
|
|
On Jun 17, 12:35 am, Klemen <klemen.zak...@gmail.com> wrote:
> Hello all,
>
> I have a variable to be explained with 12 attributes (at the moment
> all of them being arrays of 300 by 100, but in the future it can be
> even larger than 1000 by 1000 array). I would like to preform a moving
> window analysis. For each window I want to estimate the optimal
> multiple linear regression parameters.
>
> Optimal means that it is not necessary that all of 12 variables are
> used (some of them are correlated). So I want to get out an equation
> that has 2,3,4... parameters and provide the best statistics.
>
> The problem is also that not all the values within a moving window can
> be used - some pixels contain no data. Defining NaN for these values
> and using the code similar to the one written by Bringfried Stecklum
> ( http://groups.google.com/group/comp.lang.idl-pvwave/browse_t hread/
> thread/17613c70b78f1ac4/6891d260db6c7c93?
> lnk=gst&q=regression#6891d260db6c7c93), I can test which of attributes
> might be significant for the regression.
>
> The question is, how to proceed. From those attributes that I know
> that they are correlated among each other, I want to use just the one
> that explains the most variability. I can somehow imagine to select
> the final 2-4 attributes that should be used for multiple regression
> without using any FOR loops. But how do I do the final step -
> estimation of multiple regression parameters without using any loops?
> Any idea?
>
> Thank you!
> Klemen
I think the problem that you mentioned is a principle component
analysis problem (PCI). For each state you have to apply this analysis
and then select the first, second or what ever variables you need to
justify the total variance. David has a very good example on it
(http://www.dfanning.com/code_tips/pca.html).
Cheers
Dave
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
