nodrogluap
commented
4 years ago Hi, thanks for checking this out. My first intuition with the much larger distance values provided by tslearn is that it is not Z-normalizing the input data (but OpenDBA does by default). I do not use tslearn, can you confirm that this is the case? If you z-normalize and are getting different results, I would then check that they are using the White-Neely step pattern for DTW by default. If this is the case, I'd be happy to investigate further with your data and your example code for both approaches. Thanks!
I was comparing the result of pair distance from OpenDBA to tslearn side to side, and find a few datasets has really different results given the same input. For example, using the UCRArchive_2018 "Rock" data set, for the first 5 arrays, without any constraints (i.e. without itakura_max_slope or sakoe_chiba_radius) the tslearn returned the pairwise distance of 0 1 2 3 4 0 0.000000 49.989252 280.619679 704.749511 747.370318 1 49.989252 0.000000 234.680443 656.096446 710.960112 2 280.619679 234.680443 0.000000 396.202309 455.455888 3 704.749511 656.096446 396.202309 0.000000 114.267849 4 747.370318 710.960112 455.455888 114.267849 0.000000
but the OpenDBA returned the distance of 0 1 2 3 4 0 0.0 10.151 31.1862 48.1643 57.02570 1 NaN 0.000 33.6761 49.2297 58.92180 2 NaN NaN 0.0000 17.9237 17.97590 3 NaN NaN NaN 0.0000 7.05957 4 NaN NaN NaN NaN 0.00000
I also compared the order (to do clustering) in which the arrays are close to a specific array, and still found the two libraries return different results.
I am fairly new to GPU programming, so hope you can have some clue about this. I can share the code for the pairwise distance from tslearn, please let me know.