Assessment of Disparity Risks in Machine Learning Pipelines

The pervasive integration of Machine Learning (ML) algorithms across societal sectors has fundamentally transformed human interaction with technology. However, this technological advancement presents significant challenges. Research consistently reveals that these algorithms frequently amplify inherent societal biases present in datasets, resulting in ML-based systems that disproportionately impact minority populations, with particularly concerning implications in domains such as education, where decisions have lasting consequences. In response, researchers have developed various metrics and frameworks to identify outcome disparities in ML pipelines. However, these approaches face inherent limitations. Model-centric metrics, calculated over specific algorithm output spaces, are constrained by their model specificity, hindering cross-model generalization. Furthermore, these metrics are typically applied in later pipeline stages, after bias introduction. While data-centric metrics effectively detect inherent dataset biases, they fail to account for ML pipeline-specific characteristics. Additionally, the field lacks methodologies for systematic assessment of how models may amplify dataset biases. We address these methodological gaps by introducing novel approaches for assessing disparity risks in early ML pipeline stages. The research presents systematic metrics and frameworks that incorporate ML pipeline contextual specificities, advancing our understanding of algorithmic fairness. We first introduce FairEd, a systematic framework for identifying, documenting, and reporting unfairness risks. This methodology enables stakeholders to comprehend unfairness risks across environmental and analytical dimensions, providing critical insights into dataset risks, fairness notion evaluations, and mitigation effectiveness while maintaining performance metrics. Building on FairEd's insights, we develop a model-agnostic risk assessment tool, FORESEE. This decision tree-based algorithm generates predictive scores for instance-level sensitivity to protected attributes. This innovation enables stakeholders to identify discriminatory samples and predict future sample risks, overcoming model-specific assessment limitations. We then examine model selection through the lens of usable information, culminating in DispaRisk's development. This tool identifies discrimination-prone datasets and model families while enhancing bias risk explainability. DispaRisk effectively bridges data-centric and model-centric disparity metrics, offering comprehensive fairness assessment. Following DispaRisk's exploration of usable information, we explore these notions in fairness-aware contexts through the AudiRep, a framework that enables comparative analysis of Learning Fair Representations (LFR) approaches, evaluating their efficacy in reducing usable sensitive attribute information within specific ML pipeline contexts and downstream applications. This dissertation contributes by introducing novel approaches that: (i) systematically evaluate disparity risks, (ii) predict discrimination risks independent of downstream classifiers, (iii) account for pipeline-specific characteristics, and (iv) enhance explainability of model-generated disparities within specific ML contexts. These contributions advance our understanding of ML system bias, providing tools for equitable algorithm development and addressing critical gaps in timely unfairness detection. This research establishes a foundation for responsible ML system deployment across domains through comprehensive early-stage fairness assessment and mitigation frameworks.

Author

• Vasquez, Jonathan

• disparity risk
• fairness
• fairness audit
• learning fair representation
• machine learning
• usable information

Identifier

hdl:1920/14453

OAI identifier oai:identifier

oai:MARS:1920/14453