Graph structure prefix injection transformer for multi-modal entity alignment

Multi-modal entity alignment aims to integrate corresponding entities across different MMKGs. However, previous studies have not adequately considered the impact of graph structural heterogeneity on EA tasks. Different MMKGs typically exhibit variations in graph structural features, leading to distinct structural representations of the same entity relationships. Additionally, the topological structure of the graph also differs. To tackle these challenges, we introduce GSIEA, the MMEA framework that integrates structural prefix injection and modality fusion. Different from other methods that directly fuse structural data with multi-modal features to perform the alignment task, GSIEA separately processes structural data and multi-modal data such as images and attributes, incorporating a prefix injection interaction module within a multi-head attention mechanism to optimize the utilization of multi-modal information and minimize the impact of graph structural differences. Additionally, GSIEA employs a convolutional enhancement module to extract fine-grained multi-modal features and computes cross-modal weights to achieve feature fusion. We conduct experimental evaluations on two public datasets, containing 12,846 and 11,199 entity pairs, respectively, demonstrating that GSIEA outperforms baseline models, with an average improvement of 3.26% in MRR and a maximum gain of 12.5%. Furthermore, the average improvement in Hits@1 is 4.96%, with a maximum increase of 16.92%. The code of our model is stored at https://github.com/HubuKG/GSIEA.