Year: 2026
Type: conference-paper
Abstract: The spread of misinformation poses serious societal harms.While detection has been extensively studied, modeling user susceptibility remains a challenging yet important direction for strengthening users' resilience.In this work, we formally define the Personalized Misinformation Susceptibility Prediction (MiSP) task, predicting whether a specific user is susceptible to (i.e., believe) a particular misinformation claim.We construct benchmark datasets derived from social science surveys and socialplatform data, and design diverse evaluation settings covering different seen/unseen user-claim scenarios.Using this benchmark, we examine LLMs' ability to simulate personalized susceptibility under realistic multi-attribute user profiles, and further evaluate LM-, LLM-, and GNN-based approaches, revealing their limitations.Motivated by these findings, we propose VulneraSpotter, which integrates semantic information and behavioral patterns via heterogeneous and homogeneous GNNs, enabling effective generalization to previously unseen users and claims.Extensive experiments demonstrate that VulneraSpotter outperforms existing baselines in predicting personalized susceptibility, uncovers meaningful group-level patterns, and shows promising potential for mitigating misinformation spread.This work provides a benchmark and modeling foundation for user-centric misinformation understanding and mitigation.
Authors: Peng Qi, Zehong Yan, Wynne Hsu, Mong-Li Lee
Institution: National University of Singapore
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: bronze
Award: G1795493338
OpenAlex ID: W7224468375