search

wannesm / dtaidistance

Time series distances: Dynamic Time Warping (fast DTW implementation in C)
Other
1.09k stars 184 forks source link

Support for other distance metrics than euclidean #112

Closed dmehrab06 closed 2 years ago

dmehrab06 commented 3 years ago

Is it possible two calculate the DTW distance by other distance metric than euclidean. For example: I have two time series sequence of strings as follows:

a = ['rain', 'sunny', 'rain', 'cloudy'] b = ['sunny', 'rain', 'rain', 'cloudy']

I can convert them to numpy array by enumerating the strings: (rain = 0, sunny = 1, cloudy = 2)

a = [0,1,0,2] b = [1,0,0,2]

The way I want to interpret this is if two elements are same, the distance is 0; otherwise the distance is 1. So, the difference between 0--1, the difference between 1--2 and the difference between 0--2 all should be the same.

wannesm commented 3 years ago

It appears you want the Needleman-Wunsch algorithm instead of DTW. They are similar but NW takes any type of symbol and DTW only numerical inputs. See https://dtaidistance.readthedocs.io/en/latest/usage/sequence.html . In this toolbox, there is only a Python implementation for NW because it's very flexible (e.g. the substitution cost can be customized), so it's not particular fast.