On Sat, Jan 10, 2015 at 3:02 AM, Marko Dinic <marko.di...@nissatech.com> wrote:
> For example, mean of two sinusoids while one of them is shifted by Pi is > 0. And that's definitely not a good centroid in my case. Well, if you think that phase shifts represent small distance proportional to phase difference then the mean of two shifted vectors is the one that has intermediate phase. If you normalize away magnitude then you are just doing distances on a circle. The definition of the mean is the thing that minimizes squared distances to every point being averaged. You have to have a compatible mean definition to go with your distance definition. Depending on how much you are warping your signals, I would tend to look to an auto-encoder representation rather than clustering to get your similarity. The auto-encoder internal layer values are often fixed with and often have good local properties for averaging and squared differences.
