Welcome to the artifact documentation for our paper, E2Usd: Efficient-yet-effective Unsupervised State Detection for Multivariate Time Series, accepted by the research tracks of The Web Conference 2024.
This is currently a preliminary release of an informal draft version, which may have minor differences or issues. The official version will be released soon.
To run our code, we developed it using Python 3.9.18. You can install the remaining dependencies by executing the following command:
pip install -r requirements.txtWe evaluated E2USD on five real-world Unsupervised State Detection (USD) datasets and one synthetic dataset for comprehensive assessments. You can download these datasets from the following links:
| Dataset | Type | Download Link |
|---|---|---|
| MoCap | Real-world | download |
| ActRecTut | Real-world | download |
| PAMAP2 | Real-world | download |
| UscHad | Real-world | download |
| UcrSeg | Real-world | download |
| Synthetic | Synthetic | download |
After downloading the datasets, move them to the '\data' directory, ensuring the following directory structure:
.
├── data
│ ├── synthetic
│ │ ├── test0.csv
│ │ ├── test1.csv
│ │ ├── ...
│ ├── ActRecTut
│ │ ├── subject1_walk
│ │ │ ├── S111.dat
│ │ │ ├── ...
│ │ ├── subject2_walk
│ │ │ ├── S111.dat
│ │ │ ├── ...
│ ├── MoCap
│ │ ├── 4d
│ │ │ ├── amc_86_01.4d
│ │ │ ├── ...
│ │ ├── raw
│ │ │ ├── amc_86_01.txt
│ │ │ ├── ...
│ ├── PAMAP2
│ │ ├── Protocol
│ │ │ ├── subject101.dat
│ │ │ ├── ...
│ ├── USC-HAD
│ │ ├── Subject1
│ │ ├── Subject2
│ │ ├── ...
│ ├── UCRSEG
│ │ ├── Cane_100_2345.txt
│ │ ├── DutchFactory_24_2184.txt
│ │ ├── ...
Execute the train.py script located in the ./experiments directory. Make sure to specify the dataset you want to use for the experiment within the main function.
We provide a supplementary study between E2Usd and FLOSS, please refer to the Supplementary Study Between E2Usd and FLOSS. We recommend downloading the PDF to a local viewer for better readability.
This work leverages the Time2State implementation as its foundation.
Our gratitude extends to the authors of the following studies for making their datasets publicly available:
Special thanks are also due to Prof. Eamonn Keogh for his valuable feedback on our work.