I found this paper Using Pivots to Speed-Up k-Medoids Clustering in which authors explain how to use triangular geometry and cosine law to speed up search of new medoids in case of K-medoids.
My question is - can I use this approach if I'm using Dynamic Time Warping as distance measure? Or some similar measure (for example Euclidean distance that takes into count only features that exist in both vectors, when calculating distance between them)?
Or it can be used only in the case of Euclidean distance?
Short clarification about Dynamic Time Warping:
- Dynamic Time Warping finds optimal alignment between two signals that could be skewed, shifted, in certain manner
- Signals being compared may have different length
- It does that by creating a matrix that allows to repeat values from one of the signals certain number of times, to find the optimal alignment of signals
- When used as a distance measure you specify operation on two features of a signal (for example, distance of two components of vector could be (a[i] - b[i])^2)
This is explanation from top of my head, it should be more clearer from the link I specified.
So, if we have two vectors (signals) like
A = 1, 1, 1, 2, 2, 3
B = 1,2,3
D(A,B) = 0, because 1 will be repeated three times and 2 will be repeated two times in vector B.