ClaSP: parameter-free time series segmentation

Abstract The study of natural and human-made processes often results in long sequences temporally-ordered values, aka time series (TS). Such consist multiple states, e.g. operating modes a machine, such that state changes the observed result distribution shape measured values. Time segmentation (TSS) tries to find TS post-hoc deduce data-generating process. TSS is typically approached as an unsupervised learning problem aiming at identification segments distinguishable by some statistical property. Current algorithms for require domain-dependent hyper-parameters be set user, make assumptions about value or types detectable which limits their applicability. Common are measure segment homogeneity number change points, particularly hard tune each data set. We present ClaSP, novel, highly accurate, hyper-parameter-free domain-agnostic method TSS. ClaSP hierarchically splits into two parts. A point determined training binary classifier possible split selecting one best identifying subsequences from either partitions. learns its main model-parameters using novel bespoke algorithms. In our experimental evaluation benchmark 107 sets, we show outperforms art terms accuracy fast scalable. Furthermore, highlight properties several real-world case studies.