Hypotheses-Powered End-to-End Agents of Data Science
Akimov, Farkhad
Akimov, Farkhad
Author
Supervisor
Department
Machine Learning
Embargo End Date
2027-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
We introduce a novel end-to-end methodology for automatic hypothesis generation, testing, and refinement via collaboration between intelligent agents. Our Hypothesis Agent generates multiple potential explanations for patterns and trends in the datareferred to as hypotheses and iteratively adjusts them. It tests these hypotheses by leveraging automatically generated features, engineered from existing data features by the Feature Engineering Agent. Complementing these is the CTA Agent, which transforms refined hypotheses into actionable recommendations for users, emulating how data scientists summarize key insights. The agents use large language models (LLMs) equipped with the ability to execute Python code. Together, they form an end-to-end agentic pipeline that enables LLMs to collaboratively and autonomously solve comprehensive data science tasks. The pipeline automates critical steps such as data cleaning, preprocessing, feature engineering, and model training. To demonstrate practical application and enhance accessibility, we developed an interactive web application that allows users to upload datasets, visualize insights, and implement datadriven decision making strategies through an intuitive interface. Our unified system minimizes human intervention, while still allowing flexible interpretation of a wealth of patterns in the data, and automatically executing tests to validate which patterns indeed improve performance. Our extensive experimental evaluations on various benchmarks demonstrate that our system achieves sizeable improvement of up to 2.46% in classification tasks and up to 18.96% reduction in RMSE for regression tasks. These results highlight the potential of LLMs to advance the automation of complex data science workflows efficiently, while the web implementation provides a practical channel for stakeholders to leverage advanced analytical capabilities without specialized technical expertise.
Citation
Farkhad Akimov, “Hypotheses-Powered End-to-End Agents of Data Science,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Large Language Model (LLM), Data Science Agents, Hypotheses, Tabular data