Hi,
I think this is a bug in the
MultivariateNormalMixtureExpectationMaximization class. When I update the
code to allow 1 column in the rows it outputs a similar fit to matlab.
Here's an example of Matlab:
X = [normrnd(0, 1, 100, 1); normrnd(2, 2, 100, 1)]
GMModel = fitgmdist(X,2);
>> GMModel.mu
ans =
0.0737
3.0914
>> GMModel.ComponentProportion
ans =
0.6750 0.3250
>> GMModel.Sigma
ans(:,:,1) =
1.0505
ans(:,:,2) =
1.6593
I pasted the same X data into a test for
MultivariateNormalMixtureExpectationMaximization that had been updated to
allow data with a single column and get the following fit:
MultivariateNormalMixtureExpectationMaximization fitter
= new MultivariateNormalMixtureExpectationMaximization(data);
MixtureMultivariateNormalDistribution initialMix
= MultivariateNormalMixtureExpectationMaximization.estimate(data, 2);
fitter.fit(initialMix);
MixtureMultivariateNormalDistribution fittedMix = fitter.getFittedModel();
List<Pair<Double, MultivariateNormalDistribution>> components =
fittedMix.getComponents();
for (Pair<Double, MultivariateNormalDistribution> component : components) {
final double weight = component.getFirst();
final MultivariateNormalDistribution mvn = component.getSecond();
final double[] mean = mvn.getMeans();
final RealMatrix covMat = mvn.getCovariances();
System.out.printf("%s : %s : %s%n", weight, Arrays.toString(mean),
covMat.toString());
}
0.6420433138817465 : [0.016942587744259194] :
Array2DRowRealMatrix{{0.9929681356}}
0.3579566861182536 : [2.9152176347671754] :
Array2DRowRealMatrix{{1.8940290549}}
The numbers are close enough to indicate that the fit is valid.
I think the error has been in assuming that because you require 2
components to have a mixture model then you must have 2 columns in the
input data. However this is not true. You can fit single dimension data
with a mixture of single Gaussians.
Is this the functionality that you are expecting?
Regards,
Alex
On Mon, 4 Mar 2024 at 20:48, Craig Brautigam <cbrauti...@icr-team.com>
wrote:
> Forgive me if this comes in twice... I did not subscribe first before
> sending the message below.
>
>
> ________________________________
> From: Craig Brautigam
> Sent: Monday, March 4, 2024 1:33 PM
> To: user@commons.apache.org <user@commons.apache.org>
> Subject: MultivariateNormalMixtureExpectationMaximization only 1 dimension
>
> Hi,
>
> Full disclosure, I'm not a mathematician so I can not go into the weeds
> into the math. However I am tasked with porting some matlab code that is
> doing gaussian mixed model to java. I really want to use apache common
> math if possible. However the code that I'm porting has 1 dimension ( a
> single variable/attribute/property) that GMMs are being created from.
>
> MultivariateNormalMixtureExpectationMaximization looks to be a pretty
> close drop in replacement for the matlab functions
> https://www.mathworks.com/help/stats/fitgmdist.html<
> https://www.mathworks.com/help/stats/fitgmdist.html> andhttps://
> www.mathworks.com/help/stats/gmdistribution.html, however the constructor
> for MultivariateNormalMixtureExpectationMaximization clearly states the the
> number of columns in the double[][]data array MUST be no less thatn2
> columns. I'm completely baffled as to why this is the case if I want to
> try to fit data with 1 dimension in it. Is there a workaround I can use
> like provide a dummy column of data with all 0s to pacify the constructor?
> Is there another class I should be using?
>
> Any help would be greatly appreciated.
>
> Thx!
>
>
> ________________________________
> The information contained in this e-mail and any attachments from ICR,
> Inc. may contain confidential and/or proprietary information, and is
> intended only for the named recipient to whom it was originally addressed.
> If you are not the intended recipient, any disclosure, distribution, or
> copying of this e-mail or its attachments is strictly prohibited. If you
> have received this e-mail in error, please notify the sender immediately by
> return e-mail and permanently delete the e-mail and any attachments.
>
