Can one use a neural network classifier for which input data are networks?

Question

I'd like to run a classifier on data I have for which each data point is itself a graph/network. I have hundreds of graphs of different types (so labelled, though I'm interested in doing both unsupervised and supervised learning), but I'm having trouble finding out of this has been done before. Graph neural networks mainly seem to be used for within-graph classification. I've found only a little material on graph-level classification.

What are established approaches to this kind of problem? I thought of extracting features from the graphs, then using said features as inputs? Or flattening the adjacency matrices and using these as inputs to an NN. Both of these methods seem like they may discard too much information to work though?

Answer 1

YES

Graph neural networks (GNNs) can be used to make predictions about individual nodes within a graph. GNNs also can be used to make predictions about entire graphs themselves. Below, I give references to GNN use in making predictions based on molecular structures and images. You may be interested in consulting material referenced by these papers, too.

Molecules

Shui, Zeren, and George Karypis. "Heterogeneous molecular graph neural networks for predicting molecule properties." 2020 IEEE International Conference on Data Mining (ICDM). IEEE, 2020.

Wieder, Oliver, et al. "A compact review of molecular property prediction with graph neural networks." Drug Discovery Today: Technologies 37 (2020): 1-12.

Images

Hu, Haojie, et al. "Graph neural network via edge convolution for hyperspectral image classification." IEEE Geoscience and Remote Sensing Letters 19 (2021): 1-5.

Vasudevan, Varun, et al. "Image classification using graph neural network and multiscale wavelet superpixels." Pattern Recognition Letters 166 (2023): 89-96.