BSRL: Balance-guided structure refinement learning for robust link polarity prediction

Signed Graph Neural Networks (SGNNs) have demonstrated great potential in modeling trust and antagonistic relationships in online social platforms. However, real-world signed data is typically sparse and noisy, making standard SGNNs prone to overfitting unreliable local patterns. Existing robust learning methods, such as the Multiscale Social Balance (MSB) framework, mainly enhance supervision by assigning social-balance-based pseudo-labels to unlabeled edges and reweighting samples, failing to explicitly repair the corrupted underlying graph structure. To bridge this gap, we propose BSRL (Balance-Guided Structure Refinement Learning), a general and model-agnostic framework designed to boost link polarity prediction via structure optimization. The core idea of BSRL is to decouple structure refinement from representation learning. Specifically, the Balance-Guided module first utilizes structural balance theory as a prior to expand missing links, while the Structure Refinement component subsequently employs a learnable soft-pruning mechanism to filter out the noise introduced during expansion. Extensive experiments on four benchmark datasets validate the generality of BSRL, showing consistent performance gains across different backbones (SGCN and SDGNN). Notably, BSRL demonstrates superior robustness in challenging scenarios with 20% label noise: when using SDGNN as the backbone, BSRL achieves absolute Macro-F1 gains of 9.4%, 6.3%, 5.8%, and 4.5% on the Slashdot, Bitcoin-Alpha, Bitcoin-OTC, and Wiki datasets, respectively, compared to the state-of-the-art MSB method. These results indicate that explicit graph structure refinement is superior to mere label reweighting strategies when dealing with sparse and noisy data.