Abstract
We present graph partition neural networks (GPNN), an extension of graphneural networks (GNNs) able to handle extremely large graphs. GPNNs alternatebetween locally propagating information between nodes in small subgraphs andglobally propagating information between the subgraphs. To efficientlypartition graphs, we experiment with several partitioning algorithms and alsopropose a novel variant for fast processing of large scale graphs. Weextensively test our model on a variety of semi-supervised node classificationtasks. Experimental results indicate that GPNNs are either superior orcomparable to state-of-the-art methods on a wide variety of datasets forgraph-based semi-supervised classification. We also show that GPNNs can achievesimilar performance as standard GNNs with fewer propagation steps.