Context-Prompt-Align: An Architecture For Few Posts Based Social Bots Detection

In today's opinion-driven world of social media, detecting social bots is crucial. However, current methods often rely on complex, deep models that present deployment challenges and computational burdens. In our paper, we propose a three-step approach: context, prompt, and align. Context involves applying Large Language Models (LLMs) to parse and comprehend usergenerated content, thereby gaining a nuanced understanding of social interactions. Prompt strategically utilizes LLMs with InContext Learning to distill a concise yet impactful set of user posts. These curated posts are transformed into User Personas, capturing the core behaviors and characteristics of social users. Align employs Graph NeuralNetworks (GNNs) to structure these User Personas within a heterogeneous social network graph. Our method achieves superior performance compared to state-of-the-art methods, utilizing just five user posts and constructing heterogeneous social network graphs. Furthermore, our lightweight approach minimizes noise in data features and reduces computational burdens, achieving effectiveness comparable to larger models even with smaller parameter settings. Through three experimental scenarios, we have validated the effectiveness and efficiency of our method, confirming the indispensable role of each module. The open-source code will be released at https://github.com/logpum/Context-Prompt-Align.