The study of biological or physical processes often results in long sequences of temporally-ordered values, aka time series (TS). Changes in the observed processes, e.g. as a cause of natural events or internal state changes, result in changes of the measured values. Time series segmentation (TSS) tries to find such changes in TS to deduce changes in the underlying process. TSS is typically approached as an unsupervised learning problem aiming at the identification of segments distinguishable by some statistical property. We present ClaSP, a novel and highly accurate method for TSS. ClaSP hierarchically splits a TS into two parts, where each split point is determined by training a binary TS classifier for each possible split point and selecting the one with highest accuracy, i.e., the one that is best at identifying subsequences to be from either of the partitions. In our experimental evaluation using a benchmark of 98 datasets, we show that ClaSP outperforms the state-of-the-art in terms of accuracy and is also faster than the second best method. We highlight properties of ClaSP using several real-life time series.

PDF Abstract

Datasets


Introduced in the Paper:

TSSB

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Change Point Detection TSSB ClaSP Relative Change Point Distance 0.0073 # 1

Methods


No methods listed for this paper. Add relevant methods here