Search Results for « Graph Neural Networks »

Graph theory offers a versatile language for modeling relationships by representing entities as vertices interconnected by edges [1]. While undirected graphs assume symmetric connections, directed, bidirected, and multidirected variants capture richer orientation and multiplicity semantics. Graph Neural Networks (GNNs) have rapidly advanced the field of learning on graphs, with recent extensions addressing directed edges [2–4]. In this paper, we introduce two novel GNN architectures—Bidirected GNNs and Multidirected GNNs—which inherently respect endpoint-specific directions and parallel-edge counts [5–8]. We demonstrate how these models naturally encode complex orientation patterns and provide illustrative examples to highlight their expressiveness. We anticipate that our contributions will spur further advances in graph representation learning for complex network structures.

Graph theory studies the relationships among objects by modeling them as vertices (nodes) connected by edges [1-3]. Beyond the familiar undirected graph, there also exist directed, bidirected, and multidirected graphs, each encoding richer notions of orientation. Graph Neural Networks (GNNs) have emerged as a powerful framework in artificial intelligence [4-7], and recent work has extended them to directed graphs. In this paper, we further generalize directed GNNs to the bidirected and multidirected settings, developing Bidirected Graph Neural Networks and Multidirected Graph Neural Networks. We demonstrate how these extensions naturally incorporate more complex orientation patterns and anticipate that our results will spur new advances in GNN research.