Multiobjective Optimization Approaches for Bias Mitigation in Machine Learning
Open Access
- Author:
- Kamani, Mohammad Mahdi
- Graduate Program:
- Informatics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 15, 2020
- Committee Members:
- James Z Wang, Dissertation Advisor/Co-Advisor
James Z Wang, Committee Chair/Co-Chair
C Lee Giles, Committee Member
Xiang Zhang, Committee Member
Chris Eliot Forest, Outside Member
Mehrdad Mahdavi, Dissertation Advisor/Co-Advisor
Mehrdad Mahdavi, Committee Chair/Co-Chair
Jia Li, Outside Member
Mary Beth Rosson, Program Head/Chair - Keywords:
- Bias Mitigation
Multiobjective Optimization
Pareto Efficient
Pareto Frontier
Class-imbalance
Bilevel Optimization
Fairness - Abstract:
- Achieving astounding progress and evolution during the past few decades, Artificial Intelligence (AI) is becoming omnipresent in every aspect of our lives. Nowadays, AI systems are touching human lives on so many unimaginable levels, through which they are transforming our prior way of life. Despite these tremendous achievements, some challenges hinder the effectiveness of these systems in real-world applications. Bias in learning systems is one of the greatest challenges these AI algorithms are facing, which has induced both technical and social complications. Bias could appear in a variety of types and from different sources, such as class-imbalance problem, fairness issues, noise in data or labels, spurious correlations between features and labels to name but a few. Regardless of the source or type of bias, it can degrade the quality of the learned model or become a source of discrimination against minority groups in sensitive features such as race, gender, or education. In some cases, the consequences of bias could be detrimental and affect numerous human lives. Hence, it is of paramount importance to account for these effects during the training procedure of machine learning models. There are various solutions for each of the biases individually, but there is not a framework that can be adapted to a number of these biases together. In addition, there might be a trade-off between satisfying the main goal of the learning and addressing the bias issues at the same time. Thus, we ought to find solutions that entail the optimal trade-off or compromise between these conflicting objectives, which is not always considered by prior solutions. In an attempt to mitigate the effects of bias in learning tasks, in this dissertation, two multiobjective optimization approaches are proposed. The overarching goal is to reduce different bias problems into these two frameworks and solve them using proposed gradient-based algorithms. The first approach dubbed as Targeted Data-driven Regularization employs a small well-crafted target dataset that is free of bias and uses it in a bilevel programming problem to prevent the main learning process to drift toward a biased model. The inner level in this bilevel programming is designed to learn parameters of the model, and the outer level aims at finding the optimal weights for different groups or classes in the dataset using that target dataset. An application of this approach is presented for the class-imbalance problem with both synthetic and real-world empirical results that outperform the state-of-the-art. The second approach called Pareto descent optimization designed to solve multiobjective optimizations involving both the main learning objective and the bias mitigation objectives. The gradient-based algorithm is guaranteed to converge to Pareto stationary points of the problem with optimal compromises. A novel extension of this algorithm is proposed to trace points from the Pareto frontier of the problem and can converge to a solution with desired levels of compromises between objectives. This approach is generic and can be generalized to any other multiobjective optimization. This idea is utilized for the fairness problem in both supervised and unsupervised learning settings. Empirical results for both these applications assert that they can superbly exceed the state-of-the-art performance in this domain.