Distance-based classification methods

[bnaul]

It seems that in practice the most popular methods for time series classification are distance-based, typically using 1-NN with Euclidian or dynamic time warping distance. These don't exactly fit into our machine learning approach, but perhaps it would be useful to add these types of methods for the purpose of reproducing other studies / demonstrating improvements over these baselines?

[stefanv]

We've had a few false starts on DTW, which may be a good place to begin:

• https://gist.github.com/38e0bd9d41f95b5c3b43
• https://github.com/scikit-image/scikit-image/pull/518/files