license: cc-by-nc-4.0
README
Dataset Structure
HCFSLN
├── Code
├── Dataset
├── Feature_Extraction
└── readme.txt
The dataset is organized into folders based on feature type.
Within the Dataset folder:
Dataset --> M2AD --> Audio, EDA, PPG, Video, Label.csv
- Audio: Contains spectral, temporal, and frequency-based features extracted using
librosa. - EDA: Contains skin conductance levels (SCL) and peaks.
- PPG: Includes clean PPG signals processed using
NeuroKit. - Video: Includes facial expression features such as Action Units (AUs), head pose, and eye gaze, extracted with
OpenFace. - Label.csv: Provides participant identifiers and experimental conditions.
All feature files and labels should be placed in their respective folders before running the notebooks.
Each CSV file follows a standardized naming convention, for example: P7_audio_features.csv
Here, P7 represents a participant identifier, ensuring easy correlation across modalities. For instance:
- P7_audio_features.csv contains extracted audio features for participant P7.
- P7_video_features.csv contains facial expression features for participant P7.
- Similarly for P7_eda_features.csv and P7_ppg_features.csv.
Feature Extraction
Use the provided Jupyter notebooks for feature extraction:
- Audio_Features_Extraction.ipynb: Extracts audio features using
librosa. - EDA_Cleaning_Components.ipynb: Processes EDA signals using
NeuroKit. - PPG_Cleaning.ipynb: Extracts PPG signals.
Video features are extracted using OpenFace. Provide the full dataset path, and it will extract raw features from the raw video files. OpenFace repository: https://github.com/TadasBaltrusaitis/OpenFace
Ensure extracted files are saved in their designated folders inside the Feature_Extraction directory.
Code
The Code folder contains:
- Baseline models and our proposed model implementations.
- Utility scripts (utils.py), runner (runner.py), and configuration (config.py).
config.py is used to set parameters and select feature lists. Each model .py file also initializes its own parameters.
Baseline models include: cplnet.py, emotracer.py, eshms.py, hyperboliccontrastive.py, nohub.py, and tamnet.py.
All plotting and evaluation code is in utils.py. The main execution is handled by runner.py.
Important Paths to Set in config.py
ROOT_DATA = "" # Root directory where dataset folders (e.g., D1, D2, D3) are stored RESULTS_ROOT = "results" # Directory for detailed results per run (e.g., CSVs, t-SNE plots) FINAL_ROOT = "final" # Directory where aggregated final results are saved
HCFSLN Model (Our Model)
Run hcfsln.py separately to store results. Provide paths to each modality folder and label CSV separately, e.g.:
audio_folder = '/home/Audio' eda_folder = '/home/EDA' ppg_folder = '/home/PPG' video_folder = '/home/Video' labels_path = '/home/Label.csv'
Ablation Studies
For ablation experiments, set the dataset path in hcfsln_ablation.py and run it. All ablation experiments are included there.
Usage Guidelines
- Install required dependencies (librosa, NeuroKit) before processing.
- Place extracted features and labels in the correct folder structure before running models.
- Please cite relevant tools and datasets when using this work.