Understanding Belief Dynamics in Visually Interactive Explanations of Graph Neural Network Predictions

AI model explanations are intended to support users in forming and updating beliefs regarding model predictions. However, how these explanations shape cognitive belief processes remains under-explored, particularly for complex relational models like Graph Neural Networks (GNNs). We investigate how visually interactive GNN explanations influence belief dynamics through a qualitative, process-tracing think-aloud study (N=20) with domain and AI experts, complemented by descriptive summaries of coded belief processes. Participants interacted with a GNN-based healthcare satisfaction prediction model within an immersive 3D virtual reality (VR) environment. Drawing on a theoretical framework of belief dynamics, we identified how these interactive explanations shaped belief formation, prediction acceptance, and hypothesis specification. Findings reveal that while domain expertise facilitated belief formation through internal knowledge recall, AI expertise supported independent evidence search through structured reasoning strategies. Our results suggest that GNN explanations should transition from static endpoints to interactive prompts that facilitate iterative belief updating.RESEARCH HIGHLIGHTSThink-aloud study (N=20) of domain and AI experts interpreting GNN explanations in VR.Applies the belief understanding framework to trace evidence evaluation in GNN-based XAI.Contributes a reusable codebook mapping belief processes to observable interaction behaviors.GNN explanations acted as prompts for evidence search and hypothesis generation, not endpoints.Domain experts relied on local knowledge recall; AI experts on reasoning about the model.