Abstract:

The Matrix Profile (MP) of a time series has been proposed as a versatile primitive for many data mining tasks. As a companion time series, the MP records distances between nearest neighbors of sub- sequences in the original time series. The Pan Matrix Profile (PMP) is a matrix with each row being an MP corresponding to a single subsequence length and computing explicitly an exact PMP is slow. We propose an approximation algorithm called APUMPEDI to compute the PMP under the unnormalized Euclidean distance based on MP algorithms combined with interpolation. We validate their efficiency and effectiveness through extensive numerical experiments on both real-world and synthesized data sets.