2024

Nurlan Musazade; Jozsef Mezei; Xiaolu Wang

This paper investigates the impact of Large Language Models (LLMs), specifically GPT, on data analysis tasks within the framework of CRISP-DM (Cross-Industry Standard Process for Data Mining). In order to assess the efficiency of text-to-code language models in data-related tasks, we systematically examine the performance of LLMs in the stages of the data mining process. GPT models are tested against a series of Python programming and SQL tasks derived from a Master’s program’s curriculum. The tasks focus on data exploration, visualization, preprocessing, and advanced analytical tasks like association rule mining and classification. The findings show that GPT models exhibit proficiency in Python programming across various CRISP-DM stages, particularly in Data Understanding, Preparation, and Modeling. They adeptly utilize Python libraries for data manipulation and visualization, demonstrating potential as effective tools in data science. However, the study also uncovers areas where the GPT Text-to-code model shows partial correctness, highlighting the need for human oversight in complex data analysis scenarios. This research contributes to understanding how AI can augment traditional data analysis methods, particularly under the CRISP-DM framework. It reveals the potential of LLMs in automating stages of data analysis, suggesting an acceleration in analytical processes and decision-making. The study provides valuable insights for organizations integrating AI into data analysis, balancing AI strengths with human expertise.