Engineering Reliable Graph Neural Networks: A Systems Perspective
Giancarlo Manuele, Alessia Romano, Marco De Santis, Luca Ferri, Federica Moretti, Tommaso Bianchi · 2025
Graph Neural Networks (GNNs) have emerged as a transformative paradigm in machine learning, enabling effective representation learning on non-Euclidean data structures such as social networks, citation graphs, protein-interaction networks, and knowledge bases. By integrating node features and topological information through iterative message passing, GNNs have demonstrated remarkable success across a variety of tasks including node classification, link prediction, and graph-level prediction. However, the practical deployment of GNNs in real-world scenarios is hindered by several critical challenges, most notably those related to scalability, efficiency, and robustness. Scalability issues arise as the size of graphs continues to grow, with many industrial-scale networks containing millions or even billions of nodes and edges. Traditional GNN architectures suffer from the exponential growth of receptive fields, memory bottlenecks, and prohibitively high computational complexity. These limitations necessitate the development of sampling strategies, cluster-based training, and distributed systems to enable scalable learning while preserving representational fidelity. Efficiency, both at training and inference time, has become equally critical as GNNs are increasingly deployed in latency-sensitive and resourceconstrained environments. The high-dimensional matrix multiplications and irregular memory access patterns inherent in GNNs make them computationally expensive. Innovations in model simplification, quantization, pruning, and architecture-aware optimization have begun to bridge the gap between model performance and computational cost. Nevertheless, achieving hardware-aware, energy-efficient, and dynamically adaptive GNNs remains an open area of research. Robustness is another cornerstone for reliable graph-based learning. Realworld graphs are frequently incomplete, noisy, or subject to adversarial manipulation. GNNs, due to their recursive aggregation schemes, can inadvertently propagate such noise across the network, leading to degraded performance and vulnerabilities. Adversarial attacks that manipulate graph structure or node features can easily mislead predictions. Furthermore, GNNs often struggle with over-smoothing in deep architectures and lack generalization under distributional shifts. A wide range of defenses have been proposed—from adversarial training and graph purification to uncertainty modeling and certifiable robustness guarantees—but a unified and theoretically grounded approach to robust GNN design is still elusive. This survey provides an extensive and critical review of the current landscape of scalable, efficient, and robust GNNs. We organize existing methods along these three foundational axes, analyze their design principles and trade-offs, and highlight common bottlenecks. In doing so, we identify key trends, theoretical insights, and empirical practices that have shaped the development of the field. Finally, we propose open challenges and future research directions aimed at unifying scalability, efficiency, and robustness into cohesive GNN architectures suitable for real-world applications. Our goal is to offer a comprehensive reference for both researchers and practitioners looking to understand, evaluate, and build next-generation GNN systems.