Graph embeddings for temporal clustering
Take a look at the presentation.
Short summary: Predict the activity of an entity in a temporal graph stream using dynamic embeddings computed by the TGN architecture. TGN is adapted to use an RNN as a decoder which outputs the # Occurances of each entity in the datastream. The graph stream dataset is extracted from the GDELT Global Entity Graph.
I am using conda to manage the python environment. Install the environment with the provided environment.yml. Furthermore, I use sacred to manage the experiment and write results in a MongoDB.
TGN, the dynamic graph encoding architecture used in this project, is added as a git submodule.
The run.py file is the single entry point to the repo. Almost all configuration is done via the config/config.yaml, there are no parameters set in the source code. Execute python run.py print_config to see the current parameters status.
The project has multiple stages which can be (de)activated via the config file.
The data preprocessing can be executed on a cluster which uses the internal cluster memory to speed up things. After processing, the resulting database is copied back to the shared NFS.
This is not a finished project but WIP due to me leaving ScaDS.AI at the end of January 2021.
- Data preprocessing pipeline finished:
- dataset can be loaded in a specific time interval and is saved in an intermediate SQLite database
- all entity pair counts are computed and in the count per entity is calculated and stored in an extra table
- TGN is added as a submodule, however the integration with the GDELT dataset is not finished:
- add RNN decoder
- prepare input in such a form that TGN understands it
- Evaluation is still missing
You can contact me at jenspetit@posteo.net for further questions.