Information Operations (IO) on social media networks pose a significant threat to democracy and modern society, yet human detection is costly and complex. Existing supervised IO detection methods fail to capture the dynamic nature of IO user behavior, while unsupervised approaches rely on oversimplified assumptions of coordination among IO users, which may not exist in practice.
To overcome these limitations, we formulate IO user detection as an anomaly detection problem and propose a novel unsupervised approach called Temporal-bEhavior-laNguage Signals for information Operation Recognition (TENSOR).
TENSOR leverages multimodal data, including temporal online user behavior (such as message posting activities) and the textual content of messages. The motivation is that IO users typically constitute a very small fraction of all online users and exhibit unique temporal behavioral and language patterns.
Specifically, we train a Temporal Point Process (TPP) to capture abnormal temporal behavioral patterns of IO users, as they tend to behave in a coordinated manner during IO campaigns.
We further introduce a novel evidence function that converts LLM responses generated from user post timelines into quantitative scores to adjust the TPP outputs for better IO user detection. Experimental results show that TENSOR outperforms the baselines on five real-world IO datasets. Code is available at GitHub.
Blogger's Review: The TENSOR method presents an innovative solution for detecting information operations users by integrating multimodal data with time-series analysis. It not only enhances detection accuracy but also paves the way for future research, making it a worthy candidate for similar attempts in other domains.