This is the supporting website for the paper "Time Series Segmentation Applied to a New Data Set for Mobile Sensing of Human Activities". It contains the used source codes, the MOSAD data set, raw results, and analysis notebooks. It reflects the state of the paper for reproducibility and is purposely not further updated.
Human activity recognition (HAR) systems are advanced machine learning (ML) workflows that automatically detect activities from motion data, captured e.g. by wearable devices such as smartphones. These devices contain multiple sensors that record human motion as acceleration, rotation and orientation in long time series (TS) data. As a first step, HAR methods typically partition such recordings into smaller subsequences before applying more advanced feature extraction and classification techniques. In this study, we evaluate the performance of 6 classical and recently published state-of-the-art TS segmentation (TSS) algorithms on a new large HAR benchmark of 126 TS, called MOSAD, recorded with 6 participants using recent smartphone sensors. Our results show that the ClaSP algorithm achieves significantly more accurate results compared to the other methods, scoring the best segmentations in 57 out of 126 TS. This indicates that ClaSP can be a viable solution for TSS in HAR systems. Additionally, the FLOSS algorithm also shows promising results, particularly for long TS with many segments. However, there is still scope for performance improvements and further research is needed.
You can easily load MOSAD as a pandas data frame and explore the annoatated TS with our data loader. For an in-depth data exploration, see our Jupyter notebook.
>>> from src.utils import load_mosad_dataset
>>> df_mosad = load_mosad_dataset()
>>> df_mosad.head()
...We have evaluated six TSS algorithms on MOSAD. The following table summarises the average Covering performance (higher is better) and the corresponding average ranks. More details are in the paper. The raw measurements are here and an analysis Jupyter notebook is here.
| Segmentation Algorithm | Average Accuracy | Average Rank | Wins & Ties |
|---|---|---|---|
| ClaSP | 74.0% | 2.2 | 57/126 |
| FLOSS | 66.3% | 2.9 | 36/126 |
| ESPRESSO | 61.8% | 3.4 | 17/126 |
| BinSeg | 62.9% | 3.5 | 8/126 |
| Window | 56.1% | 4.1 | 10/126 |
| BOCD | 50.2% | 4.9 | 0/126 |
This repository is structured in the following way:
benchmarkcontains the source codes used for running the six TSS competitors on MOSAD.datasetsconsists the raw recordings and annotations of MOSAD as published in the paper.experimentscontains the raw measurement results for the six competitors.figuresincludes the paper plots, generated by the Jupyter notebooks.notebooksconsists of Jupyter notebooks, used to explore MOSAD and evaluate the results.srccontains the sources codes for the six competitors and utility methods.
You can download this repository (by clicking the download button in the upper right corner). As this repository is a supporting website and not an updated library, we do not recommend to install it! Download and use MOSAD and extract or adapt code snippets of interest.
If you want reference our work in your scientific publication, we would appreciate the following citation:
@inproceedings{mosad2023,
title={Time Series Segmentation Applied to a New Data Set for Mobile Sensing of Human Activities},
author={Ermshaus, Arik and Singh, Sunita and Leser, Ulf},
booktitle={DARLI-AP@EDBT/ICDT},
year={2023}
}
The sources codes for the competitors in the benchmark come from multiple authors and projects. We list here the resources we used (and adapted) for our study:
- Evaluation Metrics (https://github.com/alan-turing-institute/TCPDBench)
- ClaSP (https://sites.google.com/view/ts-parameter-free-clasp/)
- FLOSS (https://stumpy.readthedocs.io/)
- ESPRESSO (https://github.com/cruiseresearchgroup/ESPRESSO)
- BinSeg and Window (https://centre-borelli.github.io/ruptures-docs/)
- BOCD (https://github.com/y-bar/bocd)
This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License.
