Data Science is going through a transformation. By leveraging the impressive reasoning capabilities of generative AI, we can overcome manual data analysis and perform data analysis faster than ever before.
LLMs are moving rapidly from chat windows into operational pipelines. This shift is not risk-free as they are not yet trustworthy enough. LLMs suffer from hallucinations when they fabricate facts to generate answers they don't know. There is also the alignment problem, where the models misinterpret the task they were asked to perform.
In my latest paper titled “Towards automated data analysis: A guided framework for LLM-based risk estimation”, I discuss these issues and present a framework that uses LLMs to perform risk analysis, supervised by a Human-in-the-Loop. Rather than relying on a risky single-prompt approach, I propose a four-stage structured framework, which allows a human supervisor to verify the integrity and accuracy of each stage before the model moves on to the next stage.
The four stages:
1) The model identifies entities and relations of the given dataset and suggests clustering techniques.
2) Generates the code for implementing the suggested techniques.
3) The user, or an agent, executes the code.
4) The model analyzes the produced results and generates a comprehensive report.
To demonstrate the viability of this approach, the paper includes a proof of concept applying the framework to a real-world problem. We examine risk estimation of non-technical losses on power grids. The results show that within this framework, the LLM does not act just as a simple chat-bot assistant, but as a system capable of producing full, reliable data analysis reports with some degree of automation.
