Time series motif discovery.

Introduction

This example uses a dataset from the UCR Time Series Classification/Clustering collection — qtdbsele0606, an excerpt of a two-channel ECG Holter recording. We work with the normalized fragment from points 4779 to 7779 of the heartbeat series.

A single call to SAXFactory.seriesToDiscordsAndMotifs finds motifs and discords with the jMotif library:

public static void main(String[] args) throws Exception {
  Instances tsData = readTSData();
  DiscordsAndMotifs dr = SAXFactory.seriesToDiscordsAndMotifs(
      toDoubleSeries(tsData), windowSize, alphabetSize, 2, 2);
  System.out.println(dr.toString());
}

The current implementation lives in the sax-vsm_classic repository; the original jMotif Google Code project that hosted this walkthrough has been retired.

Applied to the heartbeat segment, the output looks like this:

Motifs, as a list <frequency> [<offset1>,...,<offsetN>], from last to first:
51 at [122, 123, 124, 422, 423, 570, 571, ...
57 at [106, 107, 108, 109, 110, 257, 258, ...

Discords, as a list <distance> <offset>, from last to first:
4.214421668509215 at 2305
6.7017497715148995 at 2019

This reports two top motifs and two top discords. The first motif occurs 57 times at [106, 107, 108, 109, 110, 257, 258, ...]. Note that the method reports every occurrence, so neighboring offsets such as 106, 107, 108, 109, 110 describe the same recurring pattern under a sliding window — you will usually want to collapse each run to its first offset. The strongest discord (distance 6.70 at offset 2019) marks the most unusual subsequence in the recording.