Bias in Machine Learning Software: Why? How? What to do? (2105.12195v3)

Abstract: Increasingly, software is making autonomous decisions in case of criminal sentencing, approving credit cards, hiring employees, and so on. Some of these decisions show bias and adversely affect certain social groups (e.g. those defined by sex, race, age, marital status). Many prior works on bias mitigation take the following form: change the data or learners in multiple ways, then see if any of that improves fairness. Perhaps a better approach is to postulate root causes of bias and then applying some resolution strategy. This paper postulates that the root causes of bias are the prior decisions that affect- (a) what data was selected and (b) the labels assigned to those examples. Our Fair-SMOTE algorithm removes biased labels; and rebalances internal distributions such that based on sensitive attribute, examples are equal in both positive and negative classes. On testing, it was seen that this method was just as effective at reducing bias as prior approaches. Further, models generated via Fair-SMOTE achieve higher performance (measured in terms of recall and F1) than other state-of-the-art fairness improvement algorithms. To the best of our knowledge, measured in terms of number of analyzed learners and datasets, this study is one of the largest studies on bias mitigation yet presented in the literature.

• The paper introduces Fair-SMOTE, a novel method addressing bias in machine learning by balancing protected attributes and using legal-standard situation testing.
• Fair-SMOTE balances protected attributes and uses situation testing to identify and eliminate biased labels, validated through experiments on ten datasets.
• Experiments show Fair-SMOTE effectively reduces bias and simultaneously improves or maintains performance metrics like recall and F1 score, challenging the fairness-accuracy trade-off.

An Analytical Overview of Bias in Machine Learning Software

The paper "Bias in Machine Learning Software: Why? How? What to Do?" by Chakraborty et al. provides a comprehensive paper on bias in ML systems and proposes a novel approach—Fair-SMOTE—to mitigate such bias. Given the increasing significance of autonomous decisions made by software in critical domains like criminal justice and hiring, understanding and rectifying biases embedded in these systems is crucial.

Core Concept and Methodology

Bias in ML models can have severe societal implications, often disadvantaging certain groups identified by protected attributes such as sex, race, or age. Traditional bias mitigation methods primarily focus on manipulating data or adjusting model parameters without substantial understanding of the root causes of such biases. The authors criticize these approaches for lacking domain awareness and propose Fair-SMOTE as a more principled alternative.

Fair-SMOTE builds upon the foundational ideas of SMOTE by balancing not only the class distribution but also the distribution of protected attributes associated with each class. This ensures that every class of data is represented equally across sensitive groups, addressing direct causes of bias at the data level. Additionally, Fair-SMOTE employs situation testing to identify and eliminate biased labels, a method rooted in legal standards.

Experimental Insights

The authors conducted extensive experiments across ten datasets and three ML models (logistic regression, random forest, and support vector machines) to evaluate Fair-SMOTE. The findings confirm that Fair-SMOTE effectively reduces bias without compromising performance metrics such as recall and F1 score—often the trade-off presented by alternative approaches.

A significant assertion of the paper is the possibility of enhancing fairness and model accuracy simultaneously, challenging the notion that these objectives are mutually exclusive. The experimental results substantiate this claim, demonstrating that Fair-SMOTE not only matches but frequently exceeds the fairness improvement offered by other state-of-the-art methods like Fairway and optimized pre-processing, while maintaining high performance.

Implications and Future Directions

The implications of this paper are profound, suggesting that ensuring fairness in ML systems can be integrated into standard practices without sacrificing accuracy. For practitioners and researchers alike, Fair-SMOTE offers a viable pathway to simultaneously address ethical concerns and performance requirements in ML software.

Looking forward, the paper opens avenues for further exploration, such as extending Fair-SMOTE to multi-label classification problems and enhancing its applicability in deep learning contexts. Furthermore, understanding and mitigating bias in regression tasks remains a potential area for expansion. Such advancements can significantly contribute to building equitable and reliable AI systems across diverse applications.

Concluding Remarks

In summary, the paper effectively addresses the critical issue of bias in ML systems using a novel approach grounded in domain understanding and legal testing methods. The use of Fair-SMOTE presents a promising solution that harmonizes fairness with performance, thus paving the way for more responsible and ethical AI implementations in socially impactful domains. As researchers continue to explore and refine bias mitigation techniques, approaches like Fair-SMOTE ensure that the dialogue between technical robustness and societal impact remains a priority in AI development.