Dual-process theory-driven transparent approach for seniors to accept health misinformation detection results

Abstract

Given seniors’ concerns about the reliability of black-box models in health misinformation detection (HMID), there is a pressing need for explainable HMID methods that provide transparency and instill trust. Explainable artificial intelligence (AI) aims to foster understanding and trust in AI models by implementing transparency that embodies explainability and interpretability. However, most existing explainable HMID methods solely provide post-hoc explanations and neglect the interpretation of the models’ internal logic, which prevents seniors from satisfactorily accepting the HMID results. Therefore, this study proposes a transparent Knowledge Graph-aware Two-Stage approach (KG2S) driven by the dual-process theory. KG2S combines explainability and interpretability by utilizing knowledge graphs (KGs) in two stages: Knowledge Breadth Retrieval (KBR) and Knowledge Depth Reasoning (KDR). These stages correspond with the heuristic and analytic processes of the dual-process theory, which encompasses human information processing. In the KBR stage, we leverage rich facts in KGs to replicate heuristic distillation behaviors of humans through a novel similarity-diversity twofold filter. In the KDR stage, we employ a hierarchical attention network to emulate humans’ coarse-to-fine knowledge analyses during decision-making. Extensive experiments were conducted on two real-world datasets, along with user testing, to assess the effectiveness of the proposed model. Results demonstrated that the proposed model not only outperformed competing methods in terms of HMID accuracy but also provided persuasive detection processes and reasons. Moreover, the model showed high adaptability to new topics in HMID. Our research offers valuable insights into integrating humanistic into AI algorithms and promoting the trustworthiness of AI systems.