Multimodal hazardous materials risk graph completion: a joint optimization approach with dual channel embedding and generative adversarial network

Abstract

Multimodal hazardous chemical risk knowledge graphs are gradually becoming a critical technical foundation for industrial safety management, offering novel pathways for intelligent identification and risk prediction through their integration capabilities over multi-source heterogeneous data. However, existing multimodal hazardous chemical knowledge graph face significant challenges in practical construction, including uneven distribution across modalities and severe missingness of high-dimensional feature information. These issues lead to incomplete graph structures, negatively impacting the accuracy of knowledge reasoning and risk prediction. To address these challenges, this paper proposes a multimodal knowledge graph completion model, named HCMMKGC, integrating a dual-channel embedding mechanism and generative adversarial optimization. Specifically, the dual-channel architecture independently models high- and low-dimensional multimodal data, preserving complex structural details while improving multimodal semantic consistency. Additionally, a Generative Adversarial Network is introduced to synthesize scarce modality samples, alleviating representation bias caused by modality imbalance and thereby enhancing graph completion effectiveness and downstream reasoning performance. The experimental findings demonstrate that the HCMMKGC model exhibits strong performance on the HCKG-Text and HCKG-Visual datasets, with an MRR of 0.453 and 0.414, respectively. The model's Hit@10 values of 0.642 and 0.572 are indicative of significant improvement over the existing baseline model. These results underscore the model's superior generalization capabilities and robustness.