Abstract
Recently, Graph Neural Networks (GNNs) have demonstrated remarkable success in graph-based learning tasks. However, existing deep GNN (DGNN) architectures and Neural Architecture Search (NAS) frameworks remain limited due to tightly coupled propagation–transformation operations. This leads to exponentially increasing search spaces, computational inefficiency, and depth-induced over-smoothing. To address these challenges, a Continuous Decoupled Graph Neural Architecture Search (CDGNAS) framework is proposed in this research. The proposed CDGNAS independently models structural propagation and feature transformation as separate architectural dimensions within a continuous differentiable search space. Consequently, a hardware-aware cost constraint is employed as a bi-level multi-objective optimization strategy to jointly maximize the node classification accuracy while minimizing the computational overhead. The decoupled formulation reduces the effective search-space coupling between the propagation and transformation operations. Experiments were conducted on benchmark citation networks including Cora, Citeseer, and PubMed as well as large-scale Open Graph Benchmark (OGB) datasets. The discovered architectures achieved node classification accuracies of 89.36 ± 0.30%, 77.58 ± 0.50%, and 89.15 ± 0.20% on the Cora, Citeseer, and PubMed datasets, respectively. Furthermore, the proposed CDGNAS achieved 78.92 ± 1.20% accuracy on ogbn-arxiv and 81.76 ± 0.60% ROC-AUC on the ogbn-proteins dataset. These results indicate that the CDGNAS can improve the node classification performance under the evaluated benchmark settings.