Effectiveness of similarity measures in classification of time series data with intrinsic and extrinsic variability

Time series are hard to analyse because of their intrinsic variability which arises from the stochastic nature of the underlying process. Analysis is harder still if the underlying process is non-stationary. Further extrinsic variation may be imposed by the variability of the sampling process, e.g., by sampling at different or non-uniform time intervals. We explore the efficacy of some distance/similarity measures for time series - Euclidean (EUC), neighbourhood counting metric (NCM), dynamic time warping (DTW), longest common subsequence (LCS) and all common subsequences (ACS) - for classifying time series data with and without extrinsic variability. The similarity measures are first tested on an artificial dataset containing the trajectories of a two-dimensional dynamical system. We then analyse three real datasets - the Australian Sign Language dataset (AUSLAN) (Kadous, 2002), and the KTH (Schuldt et al., 2004) and Weizmann (Gorelick et al., 2007) video sequences of human actions.