[
https://issues.apache.org/jira/browse/MATH-296?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12759203#action_12759203
]
Romeo Palijan commented on MATH-296:
------------------------------------
Eugene, that is great news. I already tried to fix it but did not get very far.
First of all, please ignore the sample values that are supposed to be from
loessFit in the example above, they are not. I have reviewed my code and I can
not reproduce those values. Sorry for that. When I try the examples again I get
hardly any changes except for the NaN values (perhaps due to the fact that the
values are not very close together?). The testcases should be deleted/changed
to reflect this I have compiled a more appropriate testcase:
|xval|0.1|0.2|0.3|0.4|0.5|0.6|0.7|0.8|0.9|1.0|1.1|1.2|1.3|1.4|1.5|1.6|1.7|1.8|1.9|2.0|
|yval|0.47|0.48|0.55|0.56|-0.08|-0.04|-0.07|-0.07|-0.56|-0.46|-0.56|-0.52|-3.03|-3.08|-3.09|-3.04|3.54|3.46|3.36|3.35|
|result|0.46076343001050907|0.49997135715509317|0.5429801876574236|0.3191000440696382|0.17824135321363765|-0.03999999999999973|-0.0699999999999999|-0.21301261447145203|-0.3260295715150817|-0.45999999999999996|-0.560000000000002|-1.4961752602921834|-2.1317826106042093|-3.0799999999999965|-3.089999999999987|-0.6229499113916326|0.9970283360414918|3.450011252797152|3.3907474604481567|3.336774229999947|
|loessFit|0.4609008580907899|0.4985267031275769|0.5408291873645029|0.308077939113221|0.1754074466314782|-0.0419929152622151|-0.07205683451999362|-0.19661654144034335|-0.31073415047540565|-0.44615854234427305|-0.5567051673879394|-1.4972970376332615|-2.1330233520442317|-3.08|-3.0900000000000043|-0.6208045742447532|0.9823842010251099|3.449395357814414|3.389793487936696|3.3359749840089385|
|weight(see
below)|1.0|1.0|1.0|1.0|1.0|1.0|1.0|1.0|1.0|1.0|1.0|1.0|0.0|0.0|1.0|1.0|0.0|0.0|1.0|1.0|
|loessFit with
weights|0.4780117093745368|0.4924876444111407|0.48398021996388496|0.3201446090891406|0.17880078999161195|-0.003305611690841502|-0.08829472923805592|-0.20902549447499932|-0.3267558674057103|-0.45542226638134214|-0.518369970805548|-0.5365908808969353|-1.492628917733731|-2.1151751796193703|-3.09|-3.0400000000000005|-2.9899999999999998|0.15500000000000005|1.7524999999999986|3.3499999999999996|
The results do not differ very much and there are no NaN, if the values are
close together. Can this be explained by a slightly different algorithm in R?
For easy copy and paste:
double[] yval={0.47, 0.48, 0.55,
0.56,-0.08,-0.04,-0.07,-0.07,-0.56,-0.46,-0.56,-0.52,-3.03,-3.08,-3.09,-3.04,
3.54, 3.46, 3.36, 3.35};
double[] xval={0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3,
1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0};
double[] weights = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1};
LoessInterpolator li = new LoessInterpolator(0.3,4);
result = li.smooth(xval, yval);
Another related thing. A very common use of the R-function is with weights
(that is how we use it as well):
*loessFit(yValues, xValues, weights)*
where weights are the values 0 or 1, I dont know if some are using weights
between 0 and 1
If a spot has weight 0 it still gets loessfitted but does not affect his
neighbours. This way you can specifiy more relevant points and have the
complete data set interpolate around these.
In your implementation you use robustnessWeights and set them initially to 1.
Do I understand this correctly that these are actually the same weights as the
ones mentioned above?
Would it be possible to add a function with a third parameter and have the
weights provided by the user? I tried just passing them trough as a parameter
instead of initializing with 1 but it did not do the trick.
> 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
> Attachments: math-296-test.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.
