Our paper “HyperNATE: Scaling Tensor-Based Hypergraph Neural Networks Through Attention” has been published in Neural Networks, the flagship Elsevier journal of the International Neural Network Society. The work extends the tensor-based hypergraph learning program begun with t-HGSP and T-HyperGNNs, tackling the central obstacle that has kept those models from large, real-world data: scale.
Why hypergraphs
Ordinary graphs connect data two nodes at a time. But many of the relationships that matter most are inherently group relationships — a set of genes that act together, a community of co-authors on a paper, a group of sensors that fail in concert. A hypergraph captures these higher-order interactions directly, with edges that can join any number of nodes at once, instead of forcing them down into a collection of pairwise links and losing information in the process.
The tensor view — and its cost
T-HyperGNNs represented hypergraphs and their signals as tensors, using the t-product algebra to define convolution on a hypergraph without the lossy clique-expansion approximations that earlier methods relied on. That gave a clean, expressive formulation — but the tensor machinery is memory- and compute-hungry, which limited it to relatively small hypergraphs.
What HyperNATE adds
HyperNATE introduces an attention mechanism into the tensor framework so the model concentrates its computation on the most informative higher-order interactions rather than treating every hyperedge equally. The result is a hypergraph neural network that keeps the expressiveness of the tensor formulation while scaling to substantially larger hypergraphs — making tensor-based hypergraph learning practical on datasets where it previously could not run.
HyperNATE joins a growing body of work in our group on machine learning for data with complex relational structure — graph and hypergraph neural networks, geometric deep learning, and graph signal processing. For the broader research program and related publications, see the research page.