Time series classification techniques
Time series classifications can be grouped based on type of discriminatory features techniques are attempting to find \cite{Bagnall_2016}:
- Whole series: two time series compared either as a vector or by a distance measure that uses all the data
- Intervals: rather than use whole series, select one or more phase dependent intervals of the series
- Shapelets: based on finding short, phase independent, patterns (shapelets) that define class, but that can appear anywhere in series. Class is distinguished by presence or absence of one or more shapelets anywhere in whole series
- Dictionary based: classification based on histograms constructed from frequency counts of recurring patterns
- Combinations: class of algorithms based on combining two or more of the above approaches into a single classifier
- Model based: Model based algorithms fit a generative model to each series then measure similarity between series using similarity between models. Commonly proposed for tasks other than classification or as part of a larger classification scheme and are often not as competitive as other approaches (except for long series of unequal length)
- Recent benchmarking analysis found that few time series classification algorithms perform better than the Dynamic Time Warping and Rotation Forest benchmark classifiers. The best alternative (COTE) was identified as being hugely computationally expensive
- Feature alignment techniques can be used to calculate relative feature-based distance measures between dissimilar time series which can then be used in clustering or classification analysis (i.e. single value representations)
Feature alignment techniques
Illustration of Dynamic Time Warping alignment process: