4 ms·
By pure coincidence, I just had a discussion at work the other day on how to automatically place labels for clusters on tSNE plots. It looks like all I need is
by vanderZwan 9y ago
By pure coincidence, I just had a discussion at work the other day on how to automatically place labels for clusters on tSNE plots. It looks like all I need is a sensible way to convert the points to a polygon first (convex hull won't work here, for obvious reason)
- hobofan 9y agoThere are ways to get the most "typical" (I think there is a better term for that) sample of a cluster, which should usually lie in the center of the cluster. AFAIK that's the traditional way to achieve what you are trying to do.
- alexcnwy 9y agoTry compute the median point across all members of the cluster then compute the cosine similarity between each point in the cluster and the median of that cluster and use the most similar as your label point.
- vanderZwan 9y agoThe clusters on our tSNE plots tend to be stretched out, have concave bits and en up intertwined - we'd have the exact same problem outlined in this article. In fact, that was the whole reason we started discussing it, since I initially proposed just doing what you suggested
- hobofan 9y agoGood to know! I must've been lucky, since most tSNE plots I've worked with were pretty tame in that respect.
- blt 9y agoLook up Alpha Shapes.
- vanderZwan 9y agoThank you! I knew it was a matter of finding the right keywords to search for. Time to experiment a bit. EDIT: Oh, wait: > Basic alpha shapes are based on the Delaunay triangulation. At that point we might as well go for DT based approaches, don't we?