ClaSP - Parameter-free Time Series Segmentation

This website contains supporting materials for paper "ClaSP - Parameter-free Time Series Segmentation" published in DAMI'23. If you want to use ClaSP in your scientific publication or application, please use the updated and maintained claspy Python package.

Source Codes 

ClaSP Code: A Python3 implementation of our code, all 107 datasets & Jupyter Notebooks

FLOSS Code: A Python3 implementation of FLOSS

ESPRESSO Code: A Matlab implementation of ESPRESSO

Ruptures Code: A  Python3 implementation of BinSeg, PELT & Window

BOCD: A Python3 implementation of BOCD

The 107 Benchmark datasets

We wish to thank the data donors and archivists of the two benchmark sets for segmentation [1] and the UCR classification archive [2].

FLOSS Benchmark Datasets: 32 segmentation datasets (see [1]) used for the experiments

Semi-synthetic Benchmark Generator: A Jupyter Notebook to generate 75 semi-synthetic segmentation datasets from the UCR/UEA classification archive (see [2])

Dataset Description (Raw CSV, Raw Images): An overview of all 107 datasets

Results presented in the paper

We provide Jupyter Notebooks and raw scores (Covering, F1) for the experimental evaluation of the paper.

Competitors

FLOSS (Covering, F1): The scores for the determination of the arc curve threshold

BinSeg (Covering, F1): The scores for the determination of cost function & CP penalty

PELT (Covering, F1): The scores for the determination of cost function & CP penalty

Window (Covering, F1): The scores for the determination of cost function & CP penalty

Design Choices

k-NN (Covering, F1): The scores for all tested k's in the k-NN in ClaSP

Classification Score (Covering, F1): The scores for all tested classification scores in ClaSP

Ensembling (Covering, F1): The scores for the determination of the iterations in the ClaSP Ensemble

SuSS (Covering, F1): The scores for the determination of the SuSS curve threshold

Window Size (Covering, F1): The scores for all tested window size selection strategies in ClaSP

CP Score Threshold (Covering, F1): The scores for the determination of the CP threshold in ClaSP

CP Gini Gain Threshold (Covering, F1): The scores for the determination of CP gini gain threshold in ClaSP

CP Rank-Sum Test (Covering, F1): The scores for all tested max. approved p-Values for the CP Rank-Sum test in ClaSP

Change Point Validation (Covering, F1): The scores for all tested change point validation techniques in ClaSP

Segmentation

Segmentation (with unknown number of CPs) (UTSA, TSSB) (Covering, F1): The scores for ClaSP and its competitors on the general segmentation task

Segmentation (with known number of CPs) (UTSA, TSSB) (Covering, F1): The scores for ClaSP and its competitors on the constraint segmentation task

Runtime (Seconds): The runtimes for ClaSP and its competitors

Other

Complex Time Series Classifiers (Covering, F1): The errors for ClaSP with different classifiers

Segmentation Visualizations: ClaSP segmentations for all 107 time series

References

[1] Supporting Website for "Domain Agnostic Online Semantic Segmentation": https://www.cs.ucr.edu/~eamonn/FLOSS/

[2] UEA & UCR Time Series classification archive: http://www.timeseriesclassification.com