Hacker News new | past | comments | ask | show | jobs | submit login

Without looking much further into it, I can't imagine it going very wrong. My intuition says that the density function will be bounded, since you're working with a finite number of points, therefore you'll always be able to climb to a local maximum.



The problem I see is the word "near" in "climb to nearest peak". To quote the curse of dimensionality wiki page:

> The common theme of these problems is that when the dimensionality increases, the volume of the space increases so fast that the available data become sparse. This sparsity is problematic for any method that requires statistical significance.

https://en.wikipedia.org/wiki/Curse_of_dimensionality


Are you concerned that in a very high dimensional space you'd need an absurd number of points in order to produce a meaninful KDE, since otherwise the points will be too sparse? If so I think I'm with you... you'll probably just end up with a very large number of small, weak clusters.

Although I wonder if you can't account for that by simply increasing the kernel bandwidth. Perhaps not. Perhaps this is why mean-shift seems to be mostly used for computer vision and image processing, where (I guess?) the number of dimensions is low.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: