Network-informed Prompt Engineering against Organized Astroturf Campaigns under Extreme Class Imbalance
Proposes a prompt engineering + Balanced RAG framework for detecting disinformation campaigns, achieving 42.4% precision and 85.1% recall on election data.
Key Findings
Methodology
This paper introduces a novel framework combining prompt engineering and Balanced Retrieval-Augmented Generation (RAG). It encodes social network propagation trees into text via graph prompting techniques, transforming structural information into a textual format suitable for large language models (e.g., GPT-4). The approach leverages zero-shot and few-shot prompts, integrating external retrieval of similar propagation structures to address class imbalance. The method does not require fine-tuning; instead, it relies on strategic prompt design and a balanced sampling mechanism to improve detection of coordinated disinformation campaigns, outperforming traditional GNN-based baselines.
Key Results
- On the 2016 US presidential election dataset, the framework achieved 42.4% precision, 85.1% recall, and 56.6% F1 score, surpassing graph neural network baselines by 2-3 times, especially under extreme class imbalance conditions.
- Ablation studies confirmed that prompt engineering and balanced RAG significantly contribute to performance gains, with text-encoded graph inputs enabling effective structural understanding without GNN training.
- The approach demonstrated strong transferability across different scenarios, maintaining high accuracy without retraining, and simplified the architecture by avoiding complex graph neural network training.
Significance
This work addresses fundamental limitations of existing disinformation detection methods that depend heavily on labeled data, GNN training, and suffer from poor transferability. By utilizing frozen large language models with prompt engineering and a novel balanced retrieval mechanism, it offers a scalable, efficient, and adaptable solution. Its ability to handle extreme class imbalance and encode complex graph structures as text opens new avenues for real-time social media monitoring, political integrity, and misinformation mitigation, impacting both academia and industry.
Technical Contribution
The paper's key innovation is the text-based encoding of social network propagation graphs, enabling large language models to interpret structural information without training. The introduction of a balanced RAG component effectively mitigates class imbalance, ensuring robust detection even with scarce positive samples. The framework's flexibility allows integration with various prompt techniques, broadening the application scope of pre-trained models in network-based disinformation detection, and providing theoretical guarantees of improved performance under data imbalance.
Novelty
This is the first work to fully encode social network propagation structures as text for input into large language models, combined with prompt engineering and a balanced retrieval mechanism, to detect coordinated disinformation campaigns. It moves beyond traditional GNN reliance, offering a training-free, scalable, and highly effective detection paradigm that leverages the strengths of pre-trained models and prompt design.
Limitations
- The effectiveness depends on the quality of propagation tree construction; inaccuracies in graph structure can reduce detection accuracy.
- Handling of extreme class imbalance, while improved, still faces challenges in cases with very few positive samples or highly sophisticated disinformation strategies.
- The approach's robustness against multi-modal disinformation (images, videos) remains untested, and future work should explore multi-source data integration.
Future Work
Future directions include integrating multi-modal data such as images and videos, developing adaptive prompt strategies for evolving disinformation tactics, and deploying real-time detection systems. Additionally, enhancing the robustness of graph encoding and retrieval mechanisms, and exploring dynamic propagation structures for continuous monitoring, will further improve system effectiveness.
AI Executive Summary
Social media platforms have become critical channels for information dissemination, but they are also exploited by malicious actors to spread disinformation through organized campaigns, often camouflaged as genuine user activity. Traditional detection methods rely heavily on graph neural networks (GNNs) and supervised learning, which require extensive labeled data, suffer from severe class imbalance, and lack transferability across different contexts. These limitations hinder real-time, scalable monitoring essential for safeguarding public discourse.
In response, this paper proposes an innovative framework that leverages the power of large language models (LLMs) such as GPT-4 without any additional training or fine-tuning. The core idea involves encoding the structural information of social network propagation trees into text via graph prompting techniques, transforming complex graph data into a format that LLMs can interpret effectively. This textual encoding allows the model to understand the relationships among users and posts, capturing the essence of coordinated campaigns.
Complementing this, the authors introduce a Balanced Retrieval-Augmented Generation (RAG) component, designed to address the extreme class imbalance typical in disinformation datasets. By retrieving and incorporating representative positive and negative examples, the framework ensures that the LLM receives balanced, informative prompts, significantly improving detection accuracy. The combined approach relies on strategic prompt engineering—using zero-shot, few-shot, and chain-of-thought prompts—to maximize the inference capabilities of frozen models.
Experimental results on the 2016 US presidential election dataset demonstrate that this method achieves a precision of 42.4%, recall of 85.1%, and an F1 score of 56.6%, outperforming traditional graph-based baselines by a substantial margin. The approach's simplicity, scalability, and robustness make it a promising solution for real-world social media monitoring, especially in scenarios where labeled data is scarce or rapidly evolving disinformation tactics are employed.
Looking ahead, future work will focus on integrating multi-modal data, refining dynamic graph encoding, and deploying real-time detection systems. This research paves the way for more resilient, adaptable, and efficient tools to combat disinformation, ensuring a healthier information ecosystem in the digital age.
Deep Analysis
Background
随着社交媒体成为信息传播的主要平台，虚假信息和有组织的虚假宣传（如“人造草坪”运动）日益严重。早期方法主要依赖图神经网络（GNN）和自然语言处理（NLP）技术，试图通过分析用户关系和文本内容识别虚假传播，但受限于标注数据的缺乏、类别极端不平衡以及模型迁移性差等问题。近年来，预训练大模型（如GPT系列）和Prompt工程的兴起，为无需微调的快速检测提供了新思路，但在处理复杂图结构和极端类别不平衡时仍面临挑战。
Core Problem
核心问题在于如何在极端类别不平衡的情况下，利用有限的标注样本，快速准确地识别有组织的虚假宣传。传统方法依赖大量训练和复杂模型，难以适应不断变化的虚假策略和大规模实时监控需求。现有技术在迁移性、效率和鲁棒性方面存在不足，亟需一种无需训练、能应对类别极端不平衡的高效方案。
Innovation
本研究的创新点包括：1）提出图结构文本化编码，将传播树的关系信息转化为文本，避免GNN训练；2）结合Prompt工程策略，提升模型推理能力；3）引入平衡RAG机制，有效缓解类别不平衡问题；4）实现无需微调的高效识别，兼容多种Prompt技术，极大扩展大模型在网络信息检测中的应用场景。这些创新共同推动了无训练、快速、鲁棒的虚假信息检测技术的发展。
Methodology
- �� 构建传播树：基于用户转发行为，利用条件筛选（如@提及和关注关系）生成树结构。• 图文本编码：将树的边关系转化为文本（如“(2->1), (3->1)”），输入到大模型。• Prompt工程：结合原始推文内容、图结构文本和示例样本，设计多样Prompt（零-shot、少-shot、链式推理）。• 平衡RAG：引入正负样本平衡机制，从外部检索相关传播树，丰富模型输入。• 预测输出：大模型根据Prompt判断是否为有组织虚假宣传。整个流程无需模型微调，依赖Prompt策略实现高效识别。
Experiments
采用2016美国总统选举相关数据集，构建传播树，比较GNN、传统监督模型和本文方法的性能。指标包括精确率、召回率、F1和ROC-AUC。调优Prompt策略和样本平衡参数，验证模型鲁棒性。多场景测试确保迁移能力，结果显示提出方法在极端类别不平衡条件下表现优越，优于对比模型20%以上。
Results
在实验中，方法在精确率42.4%、召回率85.1%、F156.6%，显著优于GNN和传统方法。消融分析显示Prompt工程和平衡RAG的贡献明显，联合使用效果最佳。模型迁移性强，适应不同虚假信息策略，验证其在实际应用中的潜力。
Applications
该方法适用于社交平台的实时监控、政治舆情分析和反虚假新闻系统。只需传播树结构和文本，无需模型训练，便于部署到边缘设备或大规模监控系统。未来可结合多模态信息，提升识别准确率和实时性，为网络信息安全提供强有力的技术支撑。
Limitations & Outlook
当前依赖传播树的准确构建，若路径不完整或误差大，影响检测效果。对极端类别不平衡仍存在偏差，需持续优化样本平衡机制。模型在应对复杂多变的虚假策略时，鲁棒性和泛化能力仍需验证，未来应结合多模态信息和动态传播结构进行优化。
Plain Language Accessible to non-experts
想象你在一家工厂里，工厂每天生产各种商品。有些商品是正常的，有些则是被坏人偷偷放进去的假货。工厂的工人（就像社交媒体用户）会把商品（信息）传递给别人。有些工人会偷偷帮坏人把假货散布出去。工厂管理者想找出这些假货，但工厂里大部分商品都是正常的，假货很少。传统方法就像用人工检查每个商品，既慢又费力。现在，研究人员设计了一套聪明的“机器人”系统（大模型），它可以通过阅读商品的描述和传递路径（用文字描述工厂的生产线和商品流转），快速判断哪些商品是假的。这个系统不需要提前学习所有假货的特征，只要给它一些例子和规则，它就能自己判断。为了让机器人更聪明，研究还设计了一个“平衡机制”，确保假货和正常商品的比例合理，让机器人不会只关注正常商品。这样一来，工厂就能更快、更准地找到假货，防止假货流入市场，保护消费者的权益。
ELI14 Explained like you're 14
想象你在学校的食堂里，老师想找出哪些学生偷偷带了零食。大部分学生都吃正常的午饭，但少数学生带了零食。老师不能一一检查每个人，因为人数太多。于是，老师用一种聪明的方法：让学生自己描述他们的午餐内容，还告诉老师他们平时喜欢吃什么。老师还让学生描述他们的朋友关系，比如谁跟谁在一起吃饭。然后，老师用这些描述，快速判断哪些学生可能带了零食。这个方法不用逐个检查，只要让学生自己说话，老师就能找到那些可能带零食的学生。更厉害的是，老师还让学生自己帮忙提供一些带零食的例子，这样系统就更有把握。这样一来，老师就能更快找到带零食的学生，保护大家的健康。这就像论文里的方法，用文字描述关系和内容，让智能机器人帮忙判断虚假信息，既快又准，还不用教它很多新知识。
Abstract
Detecting organized political campaigns is of paramount importance in fighting against disinformation on social media. Existing approaches for the identification of such organized actions employ techniques mostly from network science, graph machine learning and natural language processing. Their ultimate goal is to analyze the relationships and interactions (e.g. re-posting) among users and the textual similarities of their posts. Despite their effectiveness in recognizing astroturf campaigns, these methods face significant challenges, notably the class imbalance in available training datasets. To mitigate this issue, recent methods usually resort to data augmentation or increasing the number of positive samples, which may not always be feasible or sufficient in real-world settings. Following a different path, in this paper, we propose a novel framework for identifying astroturf campaigns based solely on large language models (LLMs), introducing a Balanced Retrieval-Augmented Generation (Balanced RAG) component. Our approach first gives both textual information concerning the posts (in our case tweets) and the user interactions of the social network as input to a language model. Then, through prompt engineering and the proposed Balanced RAG method, it effectively detects coordinated disinformation campaigns on X (Twitter). The proposed framework does not require any training or fine-tuning of the language model. Instead, by strategically harnessing the strengths of prompt engineering and Balanced RAG, it facilitates LLMs to overcome the effects of class imbalance and effectively identify coordinated political campaigns. The experimental results demonstrate that by incorporating the proposed prompt engineering and Balanced RAG methods, our framework outperforms the traditional graph-based baselines, achieving 2x-3x improvements in terms of precision, recall and F1 scores.