The goal of the project is to learn a representation Z of a pair of a time series window X and its associated graphs G. This representation can be subsequently used to perform state classification, clustering, and so on.
We use GCN+Rnn as the encoder, TNC as the a self-supervised learning method. paper
The model are going to be trained on 2 datasets.
Python version is 3.3.6, for the environment:
pip install -r requirements.txt
First, You can create the simulated dataset using the following script:
python data/simulated_data.py
The other one is eeg brain activity data with 32 features and a static graph showing the relation between these 32 features. The data format is not ready for training yet. (Update soon)
To train the TNC GCNRnnEncoder model, simply run:
python -m tnc.training --data simulation --train
After training, state classification testing can be done by
python -m tnc.training --data simulation
The Graph embedding models are in ./tnc/graph_models.
Zhang, Y., Regol, F., Valkanas, A., & Coates, M. (2022, August). Contrastive Learning for Time Series on Dynamic Graphs. In 2022 30th European Signal Processing Conference (EUSIPCO) (pp. 742-746). IEEE
@inproceedings{zhang2022contrastive,
title={Contrastive Learning for Time Series on Dynamic Graphs},
author={Zhang, Yitian and Regol, Florence and Valkanas, Antonios and Coates, Mark},
booktitle={2022 30th European Signal Processing Conference (EUSIPCO)},
pages={742--746},
year={2022},
organization={IEEE}
}