CAREER: Causal Modeling for Data Quality and Bias Mitigation

University Of California-San Diego, La Jolla CA

Investigators

Abstract

This project presents a novel approach, inspired by database methodologies, to address the significant challenge of bias in algorithmic systems, particularly in sensitive domains such as credit scoring, medical diagnostics, predictive policing, and the criminal justice system. By recognizing that such biases often stem from the underlying data, the initiative redefines algorithmic bias as a data quality management issue. Emphasizing critical aspects of data quality management such as accuracy, completeness, and consistency, the project aims to develop methods that significantly enhance the trustworthiness and societal impact of these systems. Incorporating causal modeling with these essential data quality principles, it takes a strategic approach to identifying and addressing the root causes of algorithmic bias. This effort not only marks a significant advancement in the field of data science but also contributes substantially to national and public welfare by advocating for decision-making processes that are fair, accurate, and reliable, thereby promoting national health, prosperity, and well-being in a comprehensive manner. This plan envisions a wide-ranging dissemination of its motivation, approach, and artifacts through a diverse array of interdisciplinary colloquia, seminars, and co-curricular learning opportunities. This project addresses algorithmic bias through a fourfold approach: 1) Developing new, scalable algorithms for data repair, designed for repairing data concerning a special class of integrity constraints that can capture the statistical nuances of data used for training machine learning (ML) models. 2) Establishing a holistic data debiasing framework capable of addressing various data biases and quality issues. 3) Implementing methods to quantify uncertainty in algorithmic decision-making, particularly based on ML models, where the uncertainty stems from bias and data quality issues that cannot be fully recovered and removed due to incomplete information. 4) Lastly, the project focuses on developing methods for root-cause analysis to identify underlying issues and adaptive debiasing in dynamic data environments, incorporating proactive interventions in data processing pipelines for ongoing bias mitigation. This multifaceted strategy aims to advance the fields of data quality management, data cleaning for ML, and responsible data science, significantly enhancing the reliability, fairness, and accuracy of data-driven decision-making systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.