Joint Transfer of Model Knowledge and Fairness Over Domains Using Wasserstein Distance

Abstract

Owing to the increasing use of machine learning in our daily lives, the problem of fairness has recently become an important topic in machine learning societies. Recent studies regarding fairness in machine learning have been conducted to attempt to ensure statistical independence between individual model predictions and designated sensitive attributes. However, in reality, cases exist in which the sensitive variables of data used for learning models differ from the data upon which the model is applied. In this paper, we investigate a methodology for developing a fair classification model for data with limited or no labels, by transferring knowledge from another data domain where information is fully available. This is done by controlling the Wasserstein distances between relevant distributions. Subsequently, we obtain a fair model that could be successfully applied to two datasets with different sensitive attributes. We present theoretical results validating that our approach provably transfers both classification performance and fairness over domains. Experimental results show that our method does indeed promote fairness for the target domain, while retaining reasonable classification accuracy, and that it often outperforms comparative models in terms of joint fairness.

Mitigating Inter-Subject Brain Signal Variability FOR EEG-Based Driver Fatigue State Classification

“…Inspired by [7,9], we design a EEG-based driver fatigue state classification model that achieves invariant results to individuals, while maintaining its classification accuracy. As shown in Fig.…”

“…In this paper, we introduce a unified EEG-based classification model for subjects, motivated by recent AI works of domain adaptation [7], fairness [8,9], long-tailed recognition [10,11] developed to reduce the performance gap between different data groups. Inspired by these methods, we aim to mitigate the variability of EEG signal among individuals in the driver fatigue state classification task.…”

“…Nonetheless, most existing studies for developing subject-independent models mainly focus on achieving high accuracies rather than reducing the variation among individuals. Mitigating performance variability among different data groups: Recently, there has been rapid progress in bias mitigation to reduce the performance gap between different data groups [7,8,9,10]. Ganin et al [7] proposed an adversarial training method with GRL to reduce the gap among different data domains by training their model to discriminate and indiscriminate the target and domain labels, respectively.…”

“…Ganin et al [7] proposed an adversarial training method with GRL to reduce the gap among different data domains by training their model to discriminate and indiscriminate the target and domain labels, respectively. Yoon et al [9] proposed a bias mitigation approach to achieve fair results in terms of defined demographic groups such as gender, age, and race, by minimizing a sliced Wasserstein discrepancy between two different data groups. Accordingly, we also propose a EEG-based classification method to reduce the performance gap between individuals.…”

Fair-VQA: Fairness-Aware Visual Question Answering Through Sensitive Attribute Prediction

“…In addition, the AI recruitment system used in Amazon gives higher acceptance scores for male applicants than female applicants [1]. These social issues raise criticism of existing AI systems and the need for fairness-aware AI systems in terms of sensitive attributes, such as gender, age, and ethnicity [2]- [6]. As a result, fairness studies are performed in variety of tasks, from a simple classification on structured data [7]- [9] to Natural Language Processing (NLP) [2], face recognition [3], and Generative Adversarial Network (GAN) [4].…”