OpenGNN: Augmenting Graph Neural Networks for Open-Set Node Prediction in Complex Networks

Traditional classification approaches, grounded in closed-set frameworks, face limitations in their ability to handle novel, unseen instances that frequently arise in dynamic real-world systems. The open-set classification framework plays a crucial role in addressing this challenge. In the context of node prediction within complex networks, the task of open-set prediction for unseen or unknown nodes remains a relatively unexplored frontier. The presence of these unseen nodes, capable of disrupting network dynamics, carries significant implications across various network domains. In this study, we introduce an open-set classification paradigm using Graph Neural Networks (GNNs), presenting a novel dimension to the field. The proposed framework effectively isolates unseen nodes that do not belong to any labeled training classes. Our approach enhances the vigilance, effectiveness, and robustness of GNNs that heavily rely on graph convolutions. We experimentally validate the proposed framework using three popular real-world complex networks. Importantly, our method outperforms contemporary frameworks. The OpenGNN code for reproducing the results is available for download ( https://github.com/Netralab/OpenGNN ).