A Fusion Model to Cognize the Structure of Heterogeneous Graph for Public Safety Scenarios

Heterogeneous graphs contain complex and extensive semantic information and relations. Researchers have long been devoted to uncovering latent knowledge within them, yet their application in the context of public safety scenarios has been considered scarcely. In reality, heterogeneous graphs are susceptible to both adversarial and non-adversarial disruptions. Existing works primarily rely on idealized scenarios, where experimental environments and task objectives are relatively straightforward, making it challenging to assist decision-makers in cognition and quantification of task risks. On one hand, the semantics of heterogeneous information networks evolve with task variations, making it uneasy to construct suitable data structures for different downstream tasks. On the other hand, real-world data exists redundant, absent, and adversarial, factors that undermine the generalization capability of graph models, potentially altering the outcomes of downstream tasks. Inspired by Cognitive Learning, this paper addresses relevant works on heterogeneous graphs in the real world and proposes a cognitive fusion model. Specifically, we introduce a method to subject the structure of heterogeneous graphs to adversarial attacks, which perturbs graph models. Subsequently, we employ various strategies such as graph models, attack methodologies, and dual-attention mechanisms to construct and propel the process of cognitive learning. This empowers the overall fusion model with the capability to cognize complex data structures in public safety scenarios.