Date of Award
Doctor of Philosophy in Information Systems (PhDIS)
Business and Information Systems
Machine learning (ML) models have achieved remarkable success in various applications; however, ensuring their robustness and fairness remains a critical challenge. In this research, we present a comprehensive framework designed to evaluate and benchmark ML models through the lenses of performance, explainability, and fairness. This framework addresses the increasing need for a holistic assessment of ML models, considering not only their predictive power but also their interpretability and equitable deployment.
The proposed framework leverages a multi-faceted evaluation approach, integrating performance metrics with explainability and fairness assessments. Performance evaluation incorporates standard measures such as accuracy, precision, and recall, but extends to overall balanced error rate, overall area under the receiver operating characteristic (ROC) curve (AUC), to capture model behavior across different performance aspects. Explainability assessment employs state-of-the-art techniques to quantify the interpretability of model decisions, ensuring that model behavior can be understood and trusted by stakeholders. The fairness evaluation
examines model predictions in terms of demographic parity, equalized odds, thereby addressing concerns of bias and discrimination in the deployment of ML systems.
To demonstrate the practical utility of the framework, we apply it to a diverse set of ML algorithms across various functional domains, including finance, criminology, education, and healthcare prediction. The results showcase the importance of a balanced evaluation approach, revealing trade-offs between performance, explainability, and fairness that can inform model selection and deployment decisions. Furthermore, we provide insights into the analysis of tradeoffs in selecting the appropriate model for use cases where performance, interpretability and fairness are important.
In summary, the Performance-Explainability-Fairness Framework offers a unified methodology for evaluating and benchmarking ML models, enabling practitioners and researchers to make informed decisions about model suitability and ensuring responsible and equitable AI deployment. We believe that this framework represents a crucial step towards building trustworthy and accountable ML systems in an era where AI plays an increasingly prominent role in decision-making processes.
Chakrobartty, Shuvro, "A Performance-Explainability-Fairness Framework For Benchmarking ML Models" (2023). Masters Theses & Doctoral Dissertations. 439.