You are AllSet: A Multiset Function Framework for Hypergraph Neural Networks

Hypergraphs are used to model higher-order interactions amongst agents and\nthere exist many practically relevant instances of hypergraph datasets. To\nenable efficient processing of hypergraph-structured data, several hypergraph\nneural network platforms have been proposed for learning hypergraph properties\nand structure, with a special focus on node classification. However, almost all\nexisting methods use heuristic propagation rules and offer suboptimal\nperformance on many datasets. We propose AllSet, a new hypergraph neural\nnetwork paradigm that represents a highly general framework for (hyper)graph\nneural networks and for the first time implements hypergraph neural network\nlayers as compositions of two multiset functions that can be efficiently\nlearned for each task and each dataset. Furthermore, AllSet draws on new\nconnections between hypergraph neural networks and recent advances in deep\nlearning of multiset functions. In particular, the proposed architecture\nutilizes Deep Sets and Set Transformer architectures that allow for significant\nmodeling flexibility and offer high expressive power. To evaluate the\nperformance of AllSet, we conduct the most extensive experiments to date\ninvolving ten known benchmarking datasets and three newly curated datasets that\nrepresent significant challenges for hypergraph node classification. The\nresults demonstrate that AllSet has the unique ability to consistently either\nmatch or outperform all other hypergraph neural networks across the tested\ndatasets.\n