Gaussian Mixture Models

[chris2...@gmail.com]

Hi there,

I'm attempting to use Accord's GMM to cluster a group of 100 feature (normalized) histograms, however it always seems to return the same result when classifying. I'm not doing anything fancy, just went with the default settings and ran the Compute method on the dataset. Any suggestions? Anybody else experience this problem?

Thanks

[César]

Hi there,

After you generate the model using your histograms, does the classification always returns the same class label for those same histograms? Perhaps before using the GMM you could try using a k-means instead, and see if it also has this issue. Since the GMM uses k-means internally to initialize the models, if the k-means is not working properly then possibly this could be the cause.

Please try using k-means on your data and see if you still get this issue. If the issue persists, before learning your K-Means, call its Randomize method passing true to the useSeeding parameter. This will cause k-means to use a more clever initialization algorithm which may prevent clusters from falling into the same point.

If the k-means work after that, you can pass it to the GMM afterwards, as a starting point for the clustering.

Please see if it helps!

Best regards,

Cesar