ARO-GNN: Adaptive relation-optimized graph neural networks

Existing Graph Neural Networks (GNNs) suffer from topological noise and attribute distortion when handling complex topology-attribute interactions. To overcome these limitations, we propose Adaptive Relation-Optimized Graph Neural Networks (ARO-GNN). ARO-GNN utilizes complementary information from two different graph perspectives and dynamically adjusts the graph structure. Specifically, ARO-GNN transforms the original graph structure into a dual-graph representation. This representation captures high-order relationships and global topological features, thereby addressing the limitations inherent in traditional neighborhood-based structures. In addition, ARO-GNN employs an adaptive relation optimization mechanism that dynamically adjusts the adjacency relationships in the graph by eliminating unreliable edges while incorporating missing connections. This process mitigates topological noise and attribute distortion during message passing. Experimental results show that ARO-GNN significantly improves performance across multiple public datasets for downstream tasks.