To address bias issues, fair machine learning usually jointly optimizes two (or more) metrics aiming at predictive utility and fairness. However, the inherent under-representation of minorities in the data often makes the disparate impact of subpopulations less noticeable and difficult to deal with during learning. In this paper, we propose a novel adversarial reweighting method to address such disparate impact. To balance the data distribution between the majority and the minority groups, our approach prefers samples from the majority group that are closer to the minority group as evaluated by the Wasserstein distance. Theoretical analysis shows the effectiveness of our adversarial reweighting approach. Experiments demonstrate that our approach mitigates disparate impact without sacrificing classification accuracy, outperforming related state-of-the-art methods on image and tabular benchmark datasets. Code is available at https://github.com/zhaoxuan00707/wasserstein_reweight.

Most automated hate speech detection models rely on human annotations for training and evaluation. Logic and research indicate that people who belong to groups targeted by hate speech are better at identifying it, often due to their increased familiarity with the topic and associated hate speech terminology. However, most hate speech annotation practices overlook this issue, and hence the labels produced tend to have a reduced accuracy. In this paper, we describe an approach where the text to be annotated is supplemented with background semantics, to expose the meaning of hate speech terminology that is less likely to be known to general annotators. We test the impact of this approach by measuring change in inter-annotator agreement, before and after introducing semantics, between two groups of annotators; those who belong to the target group of hate speech, and those who are not. Our experiments show that infusing text with semantic background increases inter-annotator agreement by up to 11.3% on average, aligning the annotations from annotators who do not belong to the target groups with those from the target groups.

Hate speech detection systems may exhibit discriminatory behaviours. Research in this field has focused primarily on issues of discrimination toward the language use of minoritised communities and non-White aligned English. The interrelated issues of bias, model robustness, and disproportionate harms are weakly addressed by recent evaluation approaches, which capture them only implicitly. In this paper, we recruit a multidisciplinary group of experts to bring closer this divide between fairness and trustworthy model evaluation. Specifically, we encourage the experts to discuss not only the technical, but the social, ethical, and legal aspects of this timely issue. The discussion sheds light on critical bias facets that require careful considerations when deploying hate speech detection systems in society. Crucially, they bring clarity to different approaches for assessing, becoming aware of bias from a broader perspective, and offer valuable recommendations for future research in this field.

Many Artificial Intelligence (AI) systems rely on finding patterns in large datasets, which are prone to bias and exacerbate existing segregation and inequalities of marginalised communities. Due to their socio-technical impact, bias in AI has become a pressing issue. In this work, we investigate discrimination prevention methods on the assumption that disparities of specific populations in the training samples are reproduced or even amplified in the AI system outcomes. We aim to identify the information from vulnerable groups in the training data, uncover potential inequalities in how data capture these groups and provide additional information about them to alleviate inequalities, e.g., stereotypical and generalised views that lead to learning discriminatory associations. We develop data preprocessing techniques in automated moderation (AI systems to flag or filter online abuse) due to its substantial social implications and existing challenges common to many AI applications.

Cybersickness involves all the adverse effects that can occur during a Virtual Reality (VR) immersion, which can compromise the quality of the user experience and limit the usability, functionality and duration of use of VR systems. Standardised protocols help detect stimuli that may cause cybersickness in multiple users but do not fully discriminate which specific users experience cybersickness. Of the biometric measures used to monitor cybersickness in an individual, Heart Rate Variability (HRV) is one of the most used in previous work. However, these only considered its temporal components and did not allow for rest periods between sessions, even though these can affect users’ immersion. Our analysis addresses these limitations in that changes in HRV can measure specific levels of discomfort or ”alertness” associated with the initial cybersickness stimulus induced in the 360 videos. Primarily, our empirical results show significant differences in the frequency components of HRV in response to cybersickness stimuli. These initial measurements can compete with standard subjective assessment protocols, especially for detecting whether a subject responds to a VR immersion with cybersickness symptoms.