Time Series with Matrix Profile

Matrix Profile

If you are here, you are likely aware of what a Distance Matrix (DM) is. If not, think about those tables that used to be on maps with the distance between cities. It is widely used in TS for clustering, classification, motif search etc. But, even for modestly sized datasets, the algorithms can take months to compute it and even with speed-up techniques (i.e., indexing, lower-bounding, early abandoning) they can be, at best, one or two orders of magnitude faster.

Matrix Profile it’s like a DM but faster (much faster) to compute. Figure 1 shows a DM and a Matrix Profile. As you can see, in the Matrix Profile, as the name suggests, you see the Profile of a DM. It stores the minimum Euclidean distance of every subset of one TS (think of a Sliding Window) with another (or itself, called Self-Join). It also stores a companion vector called Profile Index, that gives us the index of each nearest neighbor.

Figure 1: Distance Matrix and its Matrix Profile.

And is this good?

Why should you care?

DM usually stores redundant information, useless for most TS applications. The Matrix Profile has a host of interesting and exploitable properties. For example, the highest point on the Matrix Profile corresponds to the TS discord, the (tied) lowest points correspond to the locations of the best TS motif pair, and the variance can be seen as a measure of the TS complexity. Moreover, the histogram of the values in the Matrix Profile is the exact answer to the TS density estimation. Particularly, it has implications for TS motif discovery, TS joins, shapelet discovery (classification), density estimation, semantic segmentation, visualization, rule discovery, clustering etc.¹

This method has the following advantages/features:

• It is exact, providing no false positives or false dismissals.
• It is simple and parameter-free. In contrast, the more general metric space APSS² algorithms require building and tuning spatial access methods and/or hash functions.
• It requires an inconsequential space overhead, just O(n) with a small constant factor.
• It is extremely scalable, and for extremely large datasets we can compute the results in an anytime fashion, allowing ultra-fast approximate solutions.
• Having computed the similarity join for a dataset, we can incrementally update it very efficiently. In many domains, this means we can effectively maintain exact joins on streaming data forever.
• It provides full joins, eliminating the need to specify a similarity threshold, which is a near impossible task in this domain.
• It is embarrassingly parallelizable, both on multicore processors and in distributed systems.

set.seed(2018)
data <- cumsum(sample(c(-1, 1), 40000, TRUE))

Further readings

All information you need to get started with Matrix Profiles is available at the UCR Matrix Profile webpage. Papers, Slides, examples, are available there.

Once become more familiar with this concept, you can start reading the TSMP R Package documentation.