Clustering evaluation with Silhouette extremely slow

I am evaluating my kmeans clustering solutions using the built-in evalclusters function with Silhouette criterion:

eva = evalclusters(data,idx,'Silhouette');

Data size is 434874x4, tested on my laptop (core i7, 8GB RAM). I have been waiting for more than an hour but it still has not completed. Is there any way to boost the speed of Silhouette evaluation in Matlab?

Thanks a lot.

[Fernando Isorna Retamino]

I have the same problem. My data size is 1726944x7 and I have been waiting for more than 2 weeks with a similar computer.

eva = evalclusters(data,'kmeans','Silhouette','KList',[1:10]);

[Stephen john]

@Fernando Isorna Retamino how you solve that?

Answer 1

[John D'Errico]

Your computer is not infinitely large or infinitely fast, unless of course, you are on TV or in the movies. Then anything is computable, merely by typing a few characters, and it will happen immediately, because who wants to wait for a few hours if they are watching a movie?

In real life, big problems take time. No matter how fast is your chosen algorithm, you can always throw more data at it than the algorithm is able to handle efficiently, no matter what is the algorithm. In real life, I don't know of anybody who does not want their code to execute more quickly.

There are a few simple choices to make here:

1. Learn patience, get some coffee and read a good book while you wait. Preferably one about MATLAB, or perhaps one about numerical methods in computing.

2. Find a seriously faster computer to use. This is never as cheap as you want.

3. Choose a different, more efficient algorithm. Different tools have different properties, but faster tools need not be always as good as others.

4. Learn enough about the methods that you can implement the algorithm EFFICIENTLY in a lower level language. Note that merely compiling your code from MATLAB is rarely going to give much of a speedup. That question gets asked on a weekly basis here.

5. Learn enough about parallel computing that you can take advantage of tools like the parallel computing TB. There are many issues here, and not all codes can be sped up.

Answer 2

[neuromechanist]

A problem with evalclusters is that it essentially runs multiple k-means (or any other clustering algortihm) one by one, not in parallel. I think Adding parallel cpability to the function increases its speed dramatically, given that parallel computing toolbox is available.

evalclusters is written in object form, and I don't know how to change for loops in object oriented setups to parfor for now. Will update my response if I figure it out.

[Stephen john]

@neuromechanist have you found any update regarding that?

[neuromechanist]

Hi Stephen,

No, Unfortunately I did not find an answer to speed up the evalclusters function.

Sorry to disappoint.

[Stephen john]

@neuromechanist any other method you used to solve this?

or change algorithm?