Graph Neural Networks (GNNs) have recently become the de facto standard for modeling relational data and gain popularity in recent years.
Treditional activision functions (like ReLU and TanH) with vanilla MLP failed to encode continues representation of data. signals Periodic Activation Functions like, sin and cosin functions, hold better frequencies representation due to smoothness of the derivatives that calculated in the optimization process (the Jacobian and Hessian matrices).
In this project you would try to apply periodic activision functions in varios GNN architectures to try and improve state of the art accuracy in tasks like node classification, graph classification and edge prediction. For more details read the links below
https://www.vincentsitzmann.com/siren/
https://neptune.ai/blog/graph-neural-network-and-some-of-gnn-applications