SciCompMod/memilio

ARMAConv Hyperparametertuning

Opened this issue · 0 comments

Enhancement description

As the GNN grid search from #1070 showed, the best performing GNN is an ARMAConv model, based on the ARMAConv layer provided by the python library Spektral https://graphneural.network/layers/convolution/#armaconv

As we can see in the documentation, there are several layer specific parameters that can be altered. In order to find the best architecture for our task of predicting the diseasy dynamic for all 400 counties we conduct experiments, chinging the model parameters of the ARMAConv model.

  • hyperparameter tuning for ARMAConv (model specific parameters)
  • tuning hyperparameters like activation function and optimizer

requirements

GPU Quadro RTX 4000
tensorflow 2.9.1
numpy 1.22 4
scikit-learn 1.5.1
spektral 1.3.1
keras 2.14.0

Checklist

  • Attached labels, especially loc:: or model:: labels.
  • Linked to project