Adversarially Robust Signed Graph Contrastive Learning from Balance Augmentation

• DOI: 10.48550/arxiv.2401.10590
• Publication Date: 2024-01-01
• Authors: Jialong Zhou; Xing Ai; Yuni Lai; Kai Zhou
• Keywords: Signed graph; Robustness

Signed graphs consist of edges and signs, which can be separated into structural information balance-related information, respectively. Existing signed graph neural networks (SGNNs) typically rely on to generate embeddings. Nevertheless, the emergence recent adversarial attacks has had a detrimental impact information. Similar how structure learning restore unsigned graphs, balance applied by improving degree poisoned graph. However, this approach encounters challenge "Irreversibility Balance-related Information" - while improves, restored may not ones originally affected attacks, resulting in poor defense effectiveness. To address challenge, we propose robust SGNN framework called Balance Augmented-Signed Graph Contrastive Learning (BA-SGCL), combines principles with augmentation techniques. Experimental results demonstrate that BA-SGCL only enhances robustness against existing but also achieves superior performance link sign prediction task across various datasets.