"Transferability of Graph Neural Networks: An Extended Graphon Analysis"Maskey, SohirGraph convolutional neural networks (GCNNs) are generalizations of standard convolutional neural networks on graph structured data. GCNNs should generalize to graphs unseen during the training procedure. For this, it is important to know whether the network is transferable, namely that the network has approximately the same repercussion on graphs that represent the same phenomenon. In our work, graphs "represent the same phenomenon" if they approximate the same graphon. Graphons are continuous limits of graphs, which allow the comparison of graphs with different sizes and topologies. In this setting, we improve current results by proving transferability for general continuous filters and general unbounded graphon shift operators. |
| http://univie.ac.at/projektservice-mathematik/e/talks/Maskey_2021-06_Online_ICCHA_2021_Maskey_Graphon_Transferability.pdf |
« back