Adaptive Sampling Strategies to Construct Equitable Training Datasets

Abstract

In domains ranging from computer vision to natural language processing, machine learning models have been shown to exhibit stark disparities, often performing worse for members of traditionally underserved groups. One factor contributing to these performance gaps is a lack of representation in the data the models are trained on. It is often unclear, however, how to operationalize representativeness in specific applications. Here we formalize the problem of creating equitable training datasets, and propose a statistical framework for addressing this problem. We consider a setting where a model builder must decide how to allocate a fixed data collection budget to gather training data from different subgroups. We then frame dataset creation as a constrained optimization problem, in which one maximizes a function of group-specific performance metrics based on (estimated) group-specific learning rates and costs per sample. This flexible approach incorporates preferences of model-builders and other stakeholders, as well as the statistical properties of the learning task. When data collection decisions are made sequentially, we show that under certain conditions this optimization problem can be efficiently solved even without prior knowledge of the learning rates. To illustrate our approach, we conduct a simulation study of polygenic risk scores on synthetic genomic data-an application domain that often suffers from non-representative data collection.We find that our adaptive sampling strategy outperforms several common data collection heuristics, including equal and proportional sampling, demonstrating the value of strategic dataset design for building equitable models. CCS Concepts: • Computing methodologies → Machine learning; Artificial intelligence; • Theory of computation → Design and analysis of algorithms.

Can Strategic Data Collection Improve the Performance of Poverty Prediction Models?

“…Inspired by literature on active learning for fairness outcomes, [18,19], we also consider adaptively constructing the training dataset by increasing the sampling weights for the points from the demographic categories with the lowest accuracy (accuracy-weighted), highest mean-squared-error (MSE-weighted), or highest exclusion/inclusion error rates (disparity-weighted). In these group-based strategies, each member of the demographic category is assigned the same weight, but each group's weight differs by group-specific model performance in each round of simulated data collection.…”

Fairness and Sequential Decision Making: Limits, Lessons, and Opportunities

“…Specifically, there are often artifacts in the data, such as correlations between attributes like race or gender and the target variable, that we do not want our predictive model to learn. One way to mitigate this is to try to balance the data set to remove these correlations within the data by adding new data points or editing existing ones (data curation) [44,46,137].…”

When More Data Lead Us Astray: Active Data Acquisition in the Presence of Label Bias

“…However, the obtained labels are still assumed to be the "gold standard". Similarly, other very recent works also incorporate the fairness notion in active learning strategy design (Abernethy et al 2020;Sharaf and Daumé III 2020;Cai et al 2022). All these new approaches inherit from classical active learning the assumption that the acquired label is a perfect match with the label of interest, but this assumption does not hold in a wide range of practical scenarios.…”