Picture: Alignment of two timeseries using DTW. Source: Wikipedia
I'm trying to extract features and learn patterns from timeseries automatically. Certain algorithms however expect the training samples to be temporally aligned. Using DTW (Dynamic-Time-Warping) - Algorithm or one of its subtypes, one could align the timeseries by warping them in the time-dimension so as to fit the Average Sequence. The individual timeseries would be of arbitrary length (50 - 10000 datapoints) whereas the maximum shift along the time axis can not be forseen (which has negative implications on the processing-time).
Regarding compute-complexity (which is O(N^2) for DTW), what alternatives or alterations could I use to make an alignmet in real-time feasible?
Any help would be appreciated. Thanks!
