Graph Neural Network model for link prediction in temporal networks.
The architecture is a conjunction of :
- Graph Convolution network : for graph embedding and reduction in model parameters.
- Graph Attention network : for enhancing quality of embeddings.
- LSTM : for processing the temporal graph at each time step.
- Training error after 40 epochs : 0.36 (criterion : binary cross entropy).
- Test AUC score : 0.53.