Evaluation of time series clustering

[jonpappalord]

It would be nice to allow the user to evaluate the quality of a clustering by providing the equivalent of the silhouette score (or related metric) for time series clustering.

[rtavenar]

Hi,

This is a very good idea! I'll do that as soon as I can. If anyone is interested in giving a hand on this, the idea would be to adapt the code from sklearn to tslearn formats (shouldn't be too hard).

[fonnesbeck]

Can scikit-learn's silhouette_score not be used directly with tslearn.metrics.soft_dtw as the metric?

[rtavenar]

Hum, a first problem we would have is that we don't expect the same format for parameter X and sklearn checks that the provided X is valid.

Something that might work would be to fool sklearn by providing a reshaped X array and a metric function that reshapes arrays back before computing their similarity. Something like:

n, sz, d = X.shape
sklearn_X = X.reshape((n, -1))
sklearn_metric = lambda x, y: metric_fun(x.reshape((sz, d)), y.reshape((sz, d)))

[rtavenar]

I just pushed an attempt to provide this functionality. Could you guys test it and give some feedback about it?

[rtavenar]

@jonpappalord @fonnesbeck
Guys, could you tell if the implemented feature matches your needs? If so, I would close this Issue.

[fonnesbeck]

I'm testing it now, but I'm using a gigantic database, so it will take a few hours to complete.

[rtavenar]

Since I have no news, I close this issue, re-open it if needed.

[ceydaakbulut95]

Hello. Good to see these comments :) I have tried what you did and it worked! Thanks!

mySeries=np.array(mySeries)
n, sz, d = mySeries.shape
sklearn_X = mySeries.reshape((n, -1))
sklearn_metric = lambda x, y: metric_fun(x.reshape((sz, d)), y.reshape((sz, d)))
km.fit_predict(sklearn_X)
score = silhouette_score(sklearn_X, km.labels_, metric='euclidean')

print('Silhouetter Score: %.3f' % score)