Variance-wise Segmentation for a Temporal-Adaptive SAX

The Symbolic Aggregate approXimation algorithm (SAX) is a very popular symbolic mapping technique for time series data, and it is widely employed in pattern identification, sequence classification, abnormality detection and other data mining research. Although SAX is a general approach which is adaptable to most data, it utilises a fixed-size sliding window in order to generate motifs (temporal shapes). When certain target phenomena (activities of interest) are manifested over differing time scales, SAXmotifs are unable to correctly account for all such targets. This paper proposes a new method named the variance-wise segmentation method which can adaptively change the size of the sliding window in a generalised SAX approach. By generating motifs with differing durations, patterns can be found for activities with similar shape but occurring over a significant altered time base. This method is tested on both artificially modified ECG data, as well as, variable tactile vibration data, with improved results compared to the original SAX formulation.