How Many Iterations to Pick Best K-Cluster?

[DavidNJ]

To pick the best set of K-Means clusters centroids for each training set, how many times should be generate the centroids to find the best? 10? 100? 1000? The higher number provides the higher probability of finding the the centroids with the lowest Eout.

Thanks,

David

[MLearning]

Originally Posted by DavidNJ

To pick the best set of K-Means clusters centroids for each training set, how many times should be generate the centroids to find the best? 10? 100? 1000? The higher number provides the higher probability of finding the the centroids with the lowest Eout.

Thanks,

David

You need a condition that stops iteration when the centroids do not change.

[DavidNJ]

Different problem If you start with different initial centroids you get different results. If you run it enough you can materially change Ein and Eout because of a better centroid result.

By itself K-means clustering doesn't guarantee an optimal result although it reaches a stable result quickly (especially with so few data points).

[MLearning]

Originally Posted by DavidNJ

Different problem If you start with different initial centroids you get different results. If you run it enough you can materially change Ein and Eout because of a better centroid result.

By itself K-means clustering doesn't guarantee an optimal result although it reaches a stable result quickly (especially with so few data points).

I agree that K-means clustering is locally optimal.

[rainbow]

It was mentioned in lectures that you in general should try different initilizations.

However, I didn't do this... I just started with K random real data points. This strategy favours the SVM model when we compare classification performance, but since we execute the experiment many times this type of bias should be diminishing.

[DavidNJ]

I'm getting a difference of two letters in the answer between using the first k-means cluster returned and testing 50 clusters per training data set.

If I can't get an official answer....those who submitted the correct answer, was it one iteration or a high number of iterations to choose the centroids for RBF?

[JohnH]

Each experiment should start with a random selection of points without regard to the data set. Lloyd's algorithm is then applied to these points to discover a set of centroids. This is unsupervised learning; i.e., the training data labels are not considered. It is not intended that one should find the optimal centroids, only that one finds some set of k-means clusters.

[DavidNJ]

The location of the centroids is fully dependent on the training data. Lloyd's algorithm doesn't find an optimal solution; K-means clustering requires multiple centroid selections to determine the best fit. This reduces the 'luck' in your initial random centroid selections.

[yaser]

Originally Posted by DavidNJ

If I can't get an official answer....those who submitted the correct answer, was it one iteration or a high number of iterations to choose the centroids for RBF?

The problem specifies that each run starts with random centers, and the only time they are revised (starting with another set of random centers) is when one of the clusters becomes empty during the iterations of Lloyd's algorithm.

There are different approaches to choosing the centers that may lead to different performance. The above approach is the one used in this problem.

[DavidNJ]

Oops...less than an hour and I have to rerun all 5...