Demographically-Informed Prediction Discrepancy Index: Early Warnings of Demographic Biases for Unlabeled Populations

Published: 05 Mar 2024, Last Modified: 05 Mar 2024Accepted by TMLREveryoneRevisionsBibTeX
Abstract: An ever-growing body of work has shown that machine learning systems can be systematically biased against certain sub-populations defined by attributes like race or gender. Data imbalance and under-representation of certain populations in the training datasets have been identified as potential causes behind this phenomenon. However, understanding whether data imbalance with respect to a specific demographic group may result in biases for a given task and model class is not simple. An approach to answering this question is to perform controlled experiments, where several models are trained with different imbalance ratios and then their performance is evaluated on the target population. However, in the absence of ground-truth annotations at deployment for an unseen population, most fairness metrics cannot be computed. In this work, we explore an alternative method to study potential bias issues based on the output discrepancy of pools of models trained on different demographic groups. Models within a pool are otherwise identical in terms of architecture, hyper-parameters, and training scheme. Our hypothesis is that the output consistency between models may serve as a proxy to anticipate biases concerning demographic groups. In other words, if models tailored to different demographic groups produce inconsistent predictions, then biases are more prone to appear in the task under analysis. We formulate the Demographically-Informed Prediction Discrepancy Index (DIPDI) and validate our hypothesis in numerical experiments using both synthetic and real-world datasets. Our work sheds light on the relationship between model output discrepancy and demographic biases and provides a means to anticipate potential bias issues in the absence of ground-truth annotations. Indeed, we show how DIPDI could provide early warnings about potential demographic biases when deploying machine learning models on new and unlabeled populations that exhibit demographic shifts.
Submission Length: Long submission (more than 12 pages of main content)
Changes Since Last Submission: Final version.
Assigned Action Editor: ~Mingming_Gong1
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Number: 1899