Local graph smoothing for link prediction against universal attack

Link prediction, a crucial research topic in complex network studies, involves estimating the likelihood of links between two nodes based on known network information. This area has garnered widespread attention due to its theoretical significance and practical applications. However, existing unsupervised graph link prediction models face challenges in handling complex graph data due to their sensitivity to noise, irregularities, and varying connectivity. They frequently encounter difficulties in capturing nuanced relationships within graph structures particularly in the face of adversarial attacks. To overcome these challenges, we propose a novel approach called Local Graph Smoothing based on GNN (LGS-GNN). The main contributions of this paper include the introduction of the LGS-GNN algorithm, a GNN-based local graph smoothing method, applied to downstream tasks such as link prediction. The paper also implements and discusses universal attack strategy on the original graph data, evaluates the feasibility of the framework through theoretical discussions. Additionally, the robustness of the GNN-based LGS model is evaluated using adversarial attack methods. Experimental results demonstrate the effectiveness of the proposed approach against imperceptible perturbations, highlighting its potential for enhancing the stability and reliability of unsupervised graph link prediction models in real-world scenarios. •Introducing LGS-GNN, a novel method for noise reduction and consistency enhancement in graph data for link prediction tasks.•Implementing three attack strategies on graph data to generate adversarial samples.•Theoretical discussions on the feasibility assessment of our framework.•Evaluating the robustness of the GNN-based LGS model using adversarial attack methods in experiments.