Effective Tool Augmented Multi-Agent Framework for Data Analysis
Xilin Zhang, Zhixin Mao, Ziwen Chen, Shen Gao · Data Intelligence · 2024
Data analysis tasks aim to provide insightful analysis for given data by incorporating background knowledge of the represented phenomenon, which in turn supports decision-making. While existing large language models (LLMs) can describe data trends, they still lag behind human data analysts in terms of integrating external knowledge and in-depth data analysis. Therefore, we propose a multi-agent data analysis framework based on LLMs. The framework decomposes the data analysis task into subtasks by employing three different agents. By empowering agents with the ability to utilize data search tools, the framework enables them to search for arbitrary relevant knowledge during the analysis process, leading to more insightful analysis. Moreover, to enhance the quality of the analysis results, we propose a multi-stage iterative optimization method that iteratively performs data analysis to form more in-depth conclusions. To validate the performance of our framework, we apply it to three real-world problems in the research development of higher education in China data. Experimental results demonstrate that our approach can achieve more insightful data analysis results compared to directly using LLMs alone.