Main article

Marcus J. Holloway
School of Computer Science and Technology, Nanjing University of Posts and Telecommunications, Nanjing, China
Yuna Takahashi
Department of Information Engineering, Fukuoka Institute of Technology, Fukuoka, Japan
Pedro Alves Ferreira
Departamento de Engenharia Informática, Instituto Politécnico de Bragança, Bragança, Portugal
Wei-Lin Chang*
Department of Computer Science and Information Engineering, National Taipei University of Technology, Taipei, Taiwan
wlchang@ntut.edu.tw

DOI: https://doi.org/10.63646/jaiaa.2024.020102

Abstract

Public health adoption is governed not only by the density of social connections but also by the credibility individuals assign to information sources within those connections — a property we term the dynamic credibility state. This paper presents a deep graph learning architecture that jointly models community-level structural topology and time-varying credibility states to predict health-related behavioral adoption under external disruptions, which we refer to as privacy shocks. The proposed framework encodes heterogeneous temporal interaction graphs comprising ordinary users and healthcare professionals through a temporal graph transformer augmented with an influence-gating mechanism that weights message propagation according to source credibility. Credibility evolves continuously as a bounded latent variable driven by interaction quality, relational context, and shock perturbations that abruptly reduce perceived reliability in affected network segments. By reweighting influence pathways at each event in response to credibility changes, the model captures the empirically observed phenomenon whereby structurally equivalent contacts differ sharply in persuasive impact depending on surrounding trust conditions. Experiments conducted on large-scale synthetic temporal networks with controlled community structure, physician authority heterogeneity, and multiple shock regimes demonstrate that the proposed approach consistently outperforms static graph convolution networks, sequence-based recurrent baselines, and temporal graph representations that lack explicit credibility modeling. Ablation analyses confirm that each architectural component — temporal encoding, credibility state propagation, shock perturbation, and influence-gating — contributes independently to predictive accuracy, with the largest gains emerging under community-targeted privacy incidents. The findings suggest that integrating dynamic credibility states into graph-based adoption models substantially improves both predictive fidelity and the interpretability of influence pathways, offering actionable guidance for community-level health communication strategy.

Article details

How to Cite

J. Holloway, M., Takahashi, Y., Ferreira, P. A., & Chang, W.-L. (2024). A Deep Graph Learning Approach to Community-Level Health Adoption Prediction with Dynamic Credibility States and Shock-Driven Pathway Reweighting. Journal of AI Analytics and Applications, 2(1), 24-45. https://doi.org/10.63646/jaiaa.2024.020102