Is there any implementation of the KMD clustering in Matlab ?

In Choose Cluster Analysis Method I do not see the KMD clustering (core) clusters.

Is there any implementation of the KMD clustering in Matlab?

The following picture, from KMD clustering: Robust generic clustering of biological data, shows a visual result from KMD:

[Sim]

@MathWorks Support Team ? :-)

[the cyclist]

I won't make this an "answer", but my best guess is no.

It is a relatively new technique, so it's not too surprising if it is not in MATLAB yet. I also searched the File Exchange, and could not find any implementations of it.

[Sim]

Thanks a lot @the cyclist!

But probably, as I have just written to @Image Analyst, it might be possible to run KMDHierarchicalClustering in Matlab, maybe by following the Call Python from MATLAB page..... Just as a temporary workaround... I do not know..

Answer 1

[Image Analyst]

Well we have dbscan which would work well for bullseye-shaped data. See attached demo db scan demo. I've also included my other clustering demos for what it's worth. (Please let me know if any don't work right out of the box.)

The fact that they have clusters broken up by color where normally you would not (like the 3rd, 4th and 5th patterns in the top row) tells me that they're just displaying two features (x and y) but they're using something additional to determine clusters, because there is no way dbscan would take that pattern in the outer ring and divide it up into 6 clusters (pink, brown, green, purple, blue, and black) instead of only 1. They must have a 3rd feature, or more, that they're not plotting.

[Sim]

Thanks a lot @Image Analyst! Yes, I was thinking the same, i.e. the reasons why, in DBSCAN, they got 6 cluster/colors in the outer ring. However, if I am not wrong, I remember that you don’t have to specify the number of clusters in DBSCAN... Therefore, could it be that they set an "epsilon" parameter in such a way that their DBSCAN found 6 clusters in the outer ring? (It is just a thought)

Btw, in my case, I was thinking to use a density-based algorithm as DBSCAN, or OPTICS, or even this KMD Core, since it looks like it performs slightly better with a lot of noise and many outliers... Indeed, in my case, I have several outliers...... but, maybe, DBSCAN might be already a good solution...

Just in case that all your attached codes/algorithms are not performing in the best way for my specific case, and as a very last resort, do you think it might be possible to run KMDHierarchicalClustering in Matlab, maybe by following the Call Python from MATLAB page ? However, it is something I have never done before :-)

Just for sake of completeness, in my case (please see the photo here below, and sorry for the low quality!) I have several filament-like elements (in other words, poly-lines, i.e. lines connected by points) and since I did not find any clustering algorithm for poly-lines I thought to cluster the points forming those poly-lines.... And I thought that a density-based algorithm could work efficiently in this case...... I would exclude k-mean (I need to know the number of clusters, and therefore I cannot set the number of clusters as initial paramete), Gaussian Mixtures (I do not know which distributions they might have)....

[Image Analyst]

No, there is no distance factor that they could have chosen that would have broken up the outer ring into 6 different clusters. For example the pink and brown, and the green and brown clusters are so close that they should be one cluster. If the distance parameter was so small that those pairs were in different clusters then the blue cluster would have been broken up into multiple clusters because there are bigger gaps in the blue than there is between the pink and brown clusters.

I think dbscan would work fine for your points as long as you get the separation parameter correct.

I don't know anything about KMD or calling Python from MATLAB.

[Sim]

ah ok, thanks a lot! :-) Yes, I agree about what you wrote :-)