This repository is the reproducibility package for the “Analysis and Impact of Training Set Size in Cross-Subject Human Activity Recognition" conference paper.

Matey-Sanz, M., Torres-Sospedra, J., González-Pérez, A., Casteleyn, S., & Granell, C. (2023, November). Analysis and Impact of Training Set Size in Cross-Subject Human Activity Recognition. In Iberoamerican Congress on Pattern Recognition (pp. 391-405). Cham: Springer Nature Switzerland.

You can explore the data and code used to obtain the results presented in the paper. To properly view the Jupyter Notebook files with their rendered figures, click on the "nbviewer" badge above.

Click the on the "Binder" badge above to open an interactive Jupyter environment with all required software installed.

Install Python 3.9, download the repository, open a command line in the root of the directory and install the required software by executing:

pip install -r requirements.txt

Install Docker for building an image based on a Dockerfile with a Jupyter environment and running a container based on the image.

Download the repository, open a command line in the root of the directory and:

1. Build the image:

docker build . --tag analysis-training-har-rp

2. Run the image:

docker run -it -p 8888:8888 analysis-training-har-rp

3. Click on the login link (or copy and paste in the browser) shown in the console to access to a Jupyter environment.

Execute a script or open the desired Jupyter Notebook (*.ipynb) file. The notebook contains the code used for the analysis and its outputs. You can execute the code to reproduce the obtained results presented in the paper.

Common files:

• Dockerfile: a recipe for the computational environment using Docker.
• requirements.txt: file with the dependencies and versions used through all the code.

Code files and directories:

• functions: contains several Python files defining functions used in the following scripts and notebooks.
• 01_data-processing.py: Python script implementing the data processing described in the paper (i.e., temporal aligment, Min-Max scaling, labelling and windowing). It reads the raw data from 01_DATA/01_RAW and stores the processed data in 01_DATA/02_CLEAN (labelled clean samples) and 01_DATA/03_WINDOWED (windows). You can run the script by executing the following command in the terminal:

  python 01_data-processing.py

• 02_incremental-loso.py: Python script implementing the incremental cross-validation described in the paper. The performance metric obtained from the training process are stored in 02_MODEL-REPORTS. You can run the script with its default parameters by executing the following command in the terminal:

python 02_incremental-loso.py

You can run the following command to obtain a description of the accepted parameters:

python 02_incremental-loso.py -h

• 03_performance-evaluation-analysis.ipynb: Jupyter Notebook containing all the code used to process and analyse the performance metrics obtained from the trained models. Contains the results presented and discussed in the paper.

Directories and data files:

• 01_DATA:
  • 01_RAW: contains a raw accelerometer and gyroscope data collected from the participants of the study.
    • sXX: directory containing the raw data obtained from the subject XX. Each directory contains a json file for each activity execution sequence and for each device. Files have the following naming convention: {subject_id}_{execution}_{sw|sp}.json.
    • subjects.csv: information (e.g., age, gender) regarding the participants of the study.
  • 02_CLEAN: contains the processed accelerometer and gyroscope data, where each sample is labelled with an associated activity (i.e., SEATED, STANDING_UP, WALKING, TURNING and SITTING_DOWN). Contains a directory for each subject, and each directory has a csv file with the labelled data corresponding to one execution. The csv files follow this format: {subject_id}_{execution}_{sw|sp}.csv.
  • 03_WINDOWED: contains the windows generated from the clean data. The windows are labelled with the SEATED, STANDING_UP, WALKING, TURNING and SITTING_DOWN activities. Contains a directory for each subject with the following files:
    • {subject_id}_{sp|sw}.npy: numpy file containing the smartphone or the smartwatch data windows.
    • {subject_id}_{sp|sw}_gt.npy: numpy file containing the grountruth (i.e., associated label) for the windows in the data file.
• 02_MODEL-REPORTS: contains files with the performance metrics obtained from the trained models.
  • sp_models.csv: performance metrics from models trained with smartphone data.
  • sw_models.csv: performance metrics from models trained with smartwatch data.

• loso_groups.csv: indicates which subjects have been used to train and evaluate each model.

• 03_RESULTS: contains the results obtained from the analyses in the 03_performance-evaluation-analysis.ipynb notebook. Contains:

  • 01_{sp|sw}-ev.pdf: plots with the overall accuracy evolution for the smartphone and smartwatch models. These correspond to the Figure 2 of the paper.
  • 02_{activity}-ev.pdf: plots with the activity-wise F1-score evolution for the smartphone and smartwatch models. These correspond to the Figure 4 of the paper.
  • 03_sp-fails-by-subject.pdf: plot with the number of times each evaluation subject obtained a F1-score == 0 in the seated, standing up and sitting down activities in the sp models. The plot corresponds with the Figure 5 of the paper.
  • 04_{sp|sw}-pairwise-table.html: tables with the results of the statistical tests performed to find significant differences in the evolution of the overall performance.
  • 04_{sp|sw}-pairwise-tests.pdf: graphical representations of the previous tables. These correspond to Figure 3 of the paper.
  • 05_{overall|activities}-comparison-table.html: tables with the results of the statistical tests performed to find significant differences in the evolution of the activity-wise performance. These tables correspond the the Table 3 of the paper.

The documents in this repository are licensed under Creative Commons Attribution 4.0 International License.

All contained code is licensed under the Apache License 2.0.

All data used in this repository is licensed under Open Data Commons Attribution License.

This work has been funded by the Spanish Ministry of Universities [grants FPU19/05352 and FPU17/03832] and by MCIN/AEI/10.13039/501100011033 [grant PID2020-120250RB-I00].