by englishm on 5/26/15, 1:04 PM with 5 comments
by nicolewhite on 5/26/15, 10:53 PM
I don't understand the author's reasoning for preferring mean shift over k-means due to k-means needing the number of clusters as an input. It seems that in mean shift, choosing the kernel bandwidth parameter is just as arbitrary in that it requires domain knowledge. In k-means there are tests[0] for choosing an appropriate k. Is there a similar strategy in mean shift for choosing an appropriate kernel bandwidth parameter?
[0] http://en.wikipedia.org/wiki/Determining_the_number_of_clust...