Mailing-List: contact issues-help@commons.apache.org; run by ezmlm
Precedence: bulk
Reply-To: issues@commons.apache.org
Message-ID: <1522387978.1254060796175.JavaMail.jira@brutus>
Date: Sun, 27 Sep 2009 07:13:16 -0700 (PDT)
From: "Romeo Palijan (JIRA)" <jira@apache.org>
To: issues@commons.apache.org
Subject: [jira] Reopened: (MATH-296) LoessInterpolator.smooth() not working
 correctly
In-Reply-To: <878072684.1253135757547.JavaMail.jira@brutus>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


     [ https://issues.apache.org/jira/browse/MATH-296?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Romeo Palijan reopened MATH-296:
--------------------------------


Eugene, Luc, thanks. Great news. I have been testing this with some bigger real-life datasets (~800 points) and the results seem to be very close to R).
There is one more problem though but I think I have already found the fix. You can still get NaN-results. 
The reason is that sumWeights can become 0 and there are a couple of divisions with sumWeights. This multiplies to the degree that the complete result set is an array of NaN.
I think (but am not sure) that this happens if you have a long series of 0-weights which is bigger then the amount of points you get through the bandwidth.

I have just added the following lines after these sumWeights divisions:
after: *double meanXSquared = sumXSquared / sumWeights;*
                
                if(sumWeights==0){
                	meanX=0;
                	meanY=0;
                	meanXY=0;
                	meanXSquared=0;
                }
That did the trick for me and the results seem accurate. I will test more during the next week with even more datasets. I hope I do not make any fundamental thinking error.

If you think this is correct please add it to the code.

> LoessInterpolator.smooth() not working correctly
> ------------------------------------------------
>
>                 Key: MATH-296
>                 URL: https://issues.apache.org/jira/browse/MATH-296
>             Project: Commons Math
>          Issue Type: Bug
>    Affects Versions: 2.0
>         Environment: Java 1.6 on Vista
>            Reporter: Romeo Palijan
>             Fix For: 2.1
>
>         Attachments: math-296-test.patch, math-296.patch
>
>
> I have been comparing LoessInterpolator.smooth output with the loessFit output from R (R-project.org, probably the most widely used loess implementation) and have had strangely different numbers. I have created a small set to test the difference and something seems to be wrong with the smooth method but I do no know what and I do not understand the code.
> *Example 1*
> |x-input: |1.5| 3.0| 6| 8| 12|13| 22| 24|28|31|
> |y-input: |3.1|6.1|3.1|2.1|1.4|5.1|5.1|6.1|7.1|7.2|
> |Output LoessInterpolator.smooth():|NaN|NaN|NaN|NaN|NaN|NaN|NaN|NaN|NaN|NaN|
> |Output from loessFit() from R: |3.191178027520974|3.0407201231474037|2.7089538903778636|2.7450823274490297|4.388011000549519|4.60078952381848|5.2988217587114805|5.867536388457898|6.7797794777879705|7.444888598397342|
> *Example 2 (same x-values, y-values just floored)*
> |x-input: |1.5| 3.0| 6| 8| 12|13| 22| 24|28|31|
> |y-input: |3|6|3|2|1|5|5|6|7|7|
> |Output LoessInterpolator.smooth(): |3|6|3|2|0.9999999999999005|5.0000000000001705|5|5.999999999999972|7|6.999999999999967|
> |Output from loessFit() from R: |3.091423927353068|2.9411521572524237|2.60967950675505|2.7421759322272248|4.382996912300442|4.646774316632562|5.225153658563424|5.768301917477015|6.637079139313073|7.270482144410326|
> As you see the output is practically the replicated y-input.
> At this point this funtionality is critical for us but I could not find any other suitable java-implementation. Help. Maybe this strange behaviour gives someone a clue?

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.