Kmean Algorithm and Cosine distance

Question

I have used K-mean Algorithm with euclidean distance to cluster my dataset, then i tried cosine distance, but the algorithm does not converge with cosine metrics (it is not stopping - iteration reach to 1000 )

any suggestion please

score 0 · Answer 1 · answered Dec 06 '16 at 22:23

0

Cosine should be okay, because it is equivalent to Euclidean on L2 normalized data. Do unless you have a programming error, it should converge.

So alternatively, you could L2 normalize your data, then use regular k-means instead og spherical k-means.

Note thst Mini-Batch by design doesn't converge, in case you are using that. If you sample batches, it cannot.

answered Dec 06 '16 at 22:23

Has QUIT--Anony-Mousse

76,138
12
138
194

I did not apply L2-norm on the data ? the data is reduced using PCA , the cosine distance is applied on Pc1, Pc2 ?? – nvayien iaziz Dec 08 '16 at 12:15
No. Never apply cosine and PCA. Usually, when cosine is appropriate, then PCA is a bad idea. – Has QUIT--Anony-Mousse Dec 08 '16 at 20:44

Kmean Algorithm and Cosine distance

1 Answers1