When: Friday, November 20, 1:00 PM
Where: Tyler 106

Abstract
As algorithmic systems become increasingly embedded in decision-making processes across various domains, the need for fairness in data science has become more urgent. Many data science practitioners, particularly those with strong technical expertise but limited experience with ethical or social considerations, lack accessible tools and guidance for integrating fairness into their workflows. My research aims to bridge that gap by developing practical, research-driven approaches that enables fairness-aware decision-making in real-world data science practices. It also contributes to the broader goal of embedding fairness into the everyday practices of data scientists by creating tools and evaluations that are intuitive and workflow-compatible. My dissertation is organized into three interconnected research projects, each addressing a different point in the data science lifecycle and collectively advancing a unified agenda, i.e making fairness actionable for practitioners. The first project involves a survey-based user study evaluating the real-world applicability of the Task Fair framework (TFF) developed in prior work to identify potential bias in the data in early model development phase, with the goal of helping them make informed decisions on whether to proceed with training the data as is, apply bias mitigation techniques before training, or fully abandon the data. The second project is a benchmark designed to assess the ability of large language model (LLM) agents to perform fairness-related machine learning tasks. This project involves constructing research problems, pre-processing datasets, implementing baseline training scripts, and computing fairness and performance metrics to serve as structured input for LLM agents. It contributes to a much-needed evaluation framework for assessing how well LLM agents can perform fair machine learning decision-making tasks across different stages of the data science lifecycle. The third project models real-world biases to generate simulated data. This controlled modeling environment enables the development of diagnostic insights and practical guidance on selecting appropriate fairness interventions based on problem structure and data characteristics. My work aims to contribute to evaluating interventions that help data scientists make fairer decisions in automated systems and address bias. In this study, we target the model development, experimentation, and validation stages of the data science lifecycle.

Bio
I am a Ph.D. candidate in Computer and Statistical Science at the University of Rhode Island, working with the Machine Learning for Social and Technical Systems (ML4STS) Lab under the supervision of Dr. Sarah Brown. My research is focused on responsible AI and fair machine learning.