Bias and Fairness in Multimodal Machine Learning: A Case Study of Automated Video Interviews

Abstract

We introduce the psychometric concepts of bias and fairness in a multimodal machine learning context assessing individuals' hireability from prerecorded video interviews. We collected interviews from 733 participants and hireability ratings from a panel of trained annotators in a simulated hiring study, and then trained interpretable machine learning models on verbal, paraverbal, and visual features extracted from the videos to investigate unimodal versus multimodal bias and fairness. Our results demonstrate that, in the absence of any bias mitigation strategy, combining multiple modalities only marginally improves prediction accuracy at the cost of increasing bias and reducing fairness compared to the least biased and most fair unimodal predictor set (verbal). We further show that gendernorming predictors only reduces gender predictability for paraverbal and visual modalities, while removing gender-biased features can achieve gender blindness, minimal bias, and fairness (for all modalities except for visual) at the cost of some prediction accuracy. Overall, the reduced-feature approach using predictors from all modalities achieved the best balance between accuracy, bias, and fairness, with the verbal modality alone performing almost as well. Our analysis highlights how optimizing model prediction accuracy in isolation and in a multimodal context may cause bias, disparate impact, and potential social harm, while a more holistic optimization approach based on accuracy, bias, and fairness can avoid these pitfalls. CCS CONCEPTS• Applied computing → Law, social and behavioral sciences;• Information systems → Multimedia and multimodal retrieval; Content analysis and feature selection; • Computing methodologies → Artificial intelligence.

A Conceptual Framework for Investigating and Mitigating Machine-Learning Measurement Bias (MLMB) in Psychological Assessment

“…For example, subgroup norming of features in ML models (i.e., z -scoring ML features separately for each subgroup) is common in computer-science applications. Past research using gender norming of ML features in the context of automated interviews did not find that it substantially reduced MLMB (Booth et al, 2021). More importantly, when ML models are used in the context of selection, one should recognize that subgroup norming (on race, sex, etc.)…”

“…In addition, it is also possible for researchers to apply different types of transformations to features (e.g., log transformation, normalization) that are subgroup-specific, or the same type of transformation but with a different mathematical function between subgroups. For example, when researchers normalize features for men and women separately before model training, if the training data for men and women have different means and variances on the features, the transformations are mathematically different between gender subgroups (Booth et al, 2021). Different subgroup transformations are regarded as a type of algorithm-training bias within the MLMB framework.…”

“…Failure to adequately address measurement bias in ML measurements can have far-reaching effects because ML measurements are often used to guide, or even automate, high-stake decisions with real-life consequences. For example, using biased automated assessments of job interviews may disproportionately exclude underrepresented (gender/racial) minorities from a job or occupation (Booth et al, 2021). In other instances, ML measurements are increasingly applied to clinical and medical diagnosis (Hosseinifard et al, 2013; Kononenko, 2001), and biased ML measurements can lead to inequalities in the provision of mental-health care.…”

Towards a standard for identifying and managing bias in artificial intelligence

“…By mechanizing human characteristics these systems can obfuscate significant uncertainty and result in harmful biases. AI-based hiring systems that claim to glean information about candidates from audio and video have been shown to increase bias in outcome decisions and may present untenable trade-offs between bias mitigation and prediction accuracy [178]. AI systems marketed as making predictions based on facial expressions often generate decisions based on biased experimental design premises [172] or spurious patterns learned by the system (e.g., shortcut learning).…”

Fair Machine Learning in Healthcare: A Review

“…For example, the visual quest answering (VQA) task [6] combines computer vision (CV) and natural language processing (NLP), and the model can answer relevant questions based on medical images and clinical notes [60]. However, multi-modal models face more serious bias and fairness issues than unimodal models, despite improvements in performance [12]. Only a few works have focused on fairness issues in multi-modality in healthcare systems [17].…”