SCAF: structure-context anchored asymmetric fusion for robust multi-modal entity alignment

Multi-modal entity alignment aims to identify equivalent entities across heterogeneous multi-modal knowledge graphs by integrating information from multiple modalities. However, real-world knowledge graphs commonly suffer from severe modality missingness. Existing approaches often rely on distribution-based modality imputation to alleviate this issue, which inevitably introduces estimation noise and exacerbates cross-modal semantic inconsistency. Moreover, pronounced modality sparsity renders prevailing uniform fusion strategies highly sensitive to unreliable features, leading to noise amplification and unstable alignment results. To address these challenges, we propose SCAF, a structure-context anchored asymmetric fusion framework for robust multi-modal entity alignment. SCAF first constructs structure-context anchors from relational neighborhood topology, inducing a cross-graph consistent semantic space without relying on semantic modality completion. Within this anchor-induced space, we perform structure-context guided modality calibration independently on each graph to align the distributions of heterogeneous modalities. Finally, SCAF adopts a structure-context guided selective fusion mechanism, where anchors act as the sole queries to selectively aggregate informative modal features while effectively suppressing noise propagation. Extensive experiments on public benchmarks demonstrate that SCAF consistently outperforms representative state-of-the-art methods. Further evaluations under modality missingness and noise injection scenarios confirm its robustness. Comprehensive ablation studies additionally verify the complementary effects of structural anchoring and selective fusion.