An Interactive Approach to Bias Mitigation in Machine Learning

作者：Hao Wang, Snehasis Mukhopadhyay, Yunyu Xiao, Shiaofen Fang · 年份：2021 · DOI：10.1109/iccicc53683.2021.9811333 · 被引用次数：11 · 研究领域：Adversarial Robustness in Machine Learning、Explainable Artificial Intelligence (XAI)、Machine Learning and Data Classification

Underrepresentation and misrepresentation of protected groups in the training data is a significant source of bias for Machine Learning (ML) algorithms, resulting in decreased confidence and trustworthiness of the generated ML models. Such bias can be mitigated by incorporating both objective as well as subjective (through human users) measures of bias, and compensating for them by means of a suitable selection algorithm over subgroups of training data. In this paper, we propose a methodology of integrating bias detection and mitigation strategies through interactive visualization of machine learning models in selected protected spaces. In this approach, a (partially generated) ML model performance is visualized and evaluated by a human user or a community of human users in terms of potential presence of bias using both objective and subjective criteria. Guided by such human feedback, the ML algorithm can implement a variety of remedial sampling strategies to mitigate the bias using an iterative human-in-the-loop approach. We also provide experimental results with a benchmark ML dataset to demonstrate that such an interactive ML approach holds considerable promise in detecting and mitigating bias in ML models.