Orange Virtual Assistant: Investigating Large Language Models’ Ability to Understand and Construct Data Mining Workflows

This thesis looks into how important workflow processes can be automated by Large Language Models (LLMs) like GPT-4 to improve user engagement with Orange, a data mining application. The research specifically focuses on three primary goals: naming workflows, describing workflow functionality, and recommending new widgets to users based on partially completed workflows. Using methods such as prompt engineering and workflow similarity analysis, this study investigates LLMs' ability to comprehend and navigate enhance directed acyclic networks that serve as the foundation for Orange operations. The findings show that workflow naming and description generating activities may be effectively completed by LLMs with high accuracy, especially when high quality examples are given. While advanced models produced encouraging results in the widget suggestion task, there are still certain issues, especially with limited existing workflow examples. The results show the strengths and weaknesses of existing LLMs in helping users in constructing and understanding data mining workflows. This thesis lays the groundwork for future advancements in workflow management and user assistance by giving insightful information about the incorporation of LLMs in a data mining setting.