The BOSS in Vector Space

This page was built in support of our paper "Scalable Time Series Classification" by Patrick Schäfer.

The paper is completely self-contained, the purpose of this webpage is to provide the 1-NN BOSS VS code, the python-codes and the raw data to readers of the paper.

Slides

BOSS VS has been presented at ECML/PKDD 2016. You can download the slides here: Keynote, Powerpoint

Trade-off between accuracy and runtimes for state of the art

Runtime of state of the art on UCR datasets

Runtime of state of the art on StarLightCurves dataset

Supporting Material

• The spreadsheet which shows the test and train accuracies of each classifier. This data was used to create the critical difference diagrams.
• A spreadsheet which shows the wallclock times of each classifier on the UCR time series datasets.
• The signed rank test that shows that the 1-NN BOSS classifier is significantly better than 1-NN DTW.
• The python code to train and test the SVM classifier.
• The python code to train and test the random forest.
• Here is the c++ source code for the BOSS VS classifier.
• An alternative implementation using JAVA: Scalable Time Series Data Analytics using SFA, BOSS and BOSS VS

References

• The BOSS model
• The Shotgun Ensemble classifier
• SFA
• Data Mining and Knowledge Discovery