What will make misinformation spread: An XAI perspective

Bo, Hongbo; Wu, Yiwen; You, Zinuo; McConville, Ryan; Hong, Jun; Liu, Weiru

Abstract

Explainable Artificial Intelligence (XAI) techniques can provide explanations of how AI systems or models make decisions, or what factors AI considers when making the decisions. Online social networks have a problem with misinformation which is known to have negative effects. In this paper, we propose to utilize XAI techniques to study what factors lead to misinformation spreading by explaining a trained graph neural network that predicts misinformation spread. However, it is difficult to achieve this with the existing XAI methods for homogeneous social networks, since the spread of misinformation is often associated with heterogeneous social networks which contain different types of nodes and relationships. This paper presents, MisInfoExplainer, an XAI pipeline for explaining the factors contributing to misinformation spread in heterogeneous social networks. Firstly, a prediction module is proposed for predicting misinformation spread by leveraging GraphSAGE with heterogeneous graph convolution. Secondly, we propose an explanation module that uses gradient-based and perturbation-based methods, to identify what makes misinformation spread by explaining the trained prediction module. Experimentally we demonstrate the superiority of MisinfoExplainer in predicting misinformation spread, and also reveal the key factors that make misinformation spread by generating a global explanation for the prediction module. Finally, we conclude that the perturbation-based approach is superior to the gradient-based approach, both in terms of qualitative analysis and quantitative measurements.

Citation

Bo, H., Wu, Y., You, Z., McConville, R., Hong, J., & Liu, W. (2023). What will make misinformation spread: An XAI perspective. In L. Longo (Ed.), Explainable Artificial Intelligence (321-337). https://doi.org/10.1007/978-3-031-44067-0_17

• Conference Name: The World Conference on eXplainable Artificial Intelligence (xAI 2023)
• Conference Location: Lisbon, Portugal
• Start Date: Jul 26, 2023
• End Date: Jul 28, 2023
• Acceptance Date: May 25, 2023
• Online Publication Date: Oct 21, 2023
• Publication Date: Oct 21, 2023
• Deposit Date: Jun 27, 2023
• Publicly Available Date: Oct 22, 2025
• Publisher: Springer Verlag (Germany)
• Volume: 1902 CCIS
• Pages: 321-337
• Series Title: Communications in Computer and Information Science
• Book Title: Explainable Artificial Intelligence
• ISBN: 9783031440663
• DOI: https://doi.org/10.1007/978-3-031-44067-0_17
• Public URL: https://uwe-repository.worktribe.com/output/10889843