Is there a generalization of cosine similarity that is robust to shifts across the compared vectors? E.g. a metric assigning high similarity to the following vectors:
[0,1,1,1,2,2,0,0] [1,1,1,2,2,0,0,0]
Asked
Active
Viewed 178 times
0
-
Can you be more specific with what "robust to shifts" means? – David Apr 03 '17 at 17:19
-
That accounts for misalignments of highly similar fractions of the compared feature vectors – Dion Apr 03 '17 at 17:20
-
Ah, as in everything was shifted one unit to the right...You may have to iteratively try some lags & leads of your data (a al cross-correlation). Obviously this can get computationally expensive, though. – David Apr 03 '17 at 17:41
-
This might be better suited for stats.stackexchange or dsp.stackexchange (signal processing) – David Apr 03 '17 at 17:42