Graph Learning with Distributional Edge Layouts

arXiv:2402.16402v1 Announce Type: new
Abstract: Graph Neural Networks (GNNs) learn from graph-structured data by passing local messages between neighboring nodes along edges on certain topological layouts. Typically, these topological layouts in modern GNNs are deterministically computed (e.g., attention-based GNNs) or locally sampled (e.g., GraphSage) under heuristic assumptions. In this paper, we for the first time pose that these layouts can be globally sampled via Langevin dynamics following Boltzmann distribution equipped with explicit physical energy, leading to higher feasibility in the …

abstract arxiv assumptions attention cs.ai cs.lg data edge gnns graph graph learning graph neural networks learn messages modern networks neural networks nodes paper structured data type