# Pro or Anti? A Social Influence Model of Online Stance Flipping

arXiv: https://arxiv.org/abs/2106.11076 · HTML: https://mlsanity.com/paper/2106.11076/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 21 Jun 2021.

## Built on by (most cited first, 6 of 6)

- [Prompting and Fine-Tuning Open-Sourced Large Language Models for Stance Classification](https://mlsanity.com/paper/2309.13734/) (2023)
- [Examining the Influence of Political Bias on Large Language Model Performance in Stance Classification](https://mlsanity.com/paper/2407.17688/) (2024)
- [Are LLM-Powered Social Media Bots Realistic?](https://mlsanity.com/paper/2508.00998/) (2025)
- [Social Theory Should Be a Structural Prior for Agentic AI: A Formal Framework for Multi-Agent Social Systems](https://mlsanity.com/paper/2605.07069/) (2026)
- [Simulating Misinformation Vulnerabilities With Agent Personas](https://mlsanity.com/paper/2511.04697/) (2025)
- [What is a Social Media Bot? A Global Comparison of Bot and Human Characteristics](https://mlsanity.com/paper/2501.00855/) (2025)

## Abstract (arXiv)

Social influence characterizes the change of an individual's stances in a complex social environment towards a topic. Two factors often govern the influence of stances in an online social network: endogenous influences driven by an individual's innate beliefs through the agent's past stances and exogenous influences formed by social network influence between users. Both endogenous and exogenous influences offer important cues to user susceptibility, thereby enhancing the predictive performance on stance changes or flipping. In this work, we propose a stance flipping prediction problem to identify Twitter agents that are susceptible to stance flipping towards the coronavirus vaccine (i.e., from pro-vaccine to anti-vaccine). Specifically, we design a social influence model where each agent has some fixed innate stance and a conviction of the stance that reflects the resistance to change; agents influence each other through the social network structure.From data collected between April 2020 to May 2021, our model achieves 86% accuracy in predicting agents that flip stances. Further analysis identifies that agents that flip stances have significantly more neighbors engaging in collective expression of the opposite stance, and 53.7% of the agents that flip stances are bots and bot agents require lesser social influence to flip stances.