Abstract:We present Shades-of-NULL, a benchmark for responsible missing value imputation. Our benchmark includes state-of-the-art imputation techniques, and embeds them into the machine learning development lifecycle. We model realistic missingness scenarios that go beyond Rubin's classic Missing Completely at Random (MCAR), Missing At Random (MAR) and Missing Not At Random (MNAR), to include multi-mechanism missingness (when different missingness patterns co-exist in the data) and missingness shift (when the missingness mechanism changes between training and test). Another key novelty of our work is that we evaluate imputers holistically, based on the predictive performance, fairness and stability of the models that are trained and tested on the data they produce. We use Shades-of-NULL to conduct a large-scale empirical study involving 20,952 experimental pipelines, and find that, while there is no single best-performing imputation approach for all missingness types, interesting performance patterns do emerge when comparing imputer performance in simpler vs. more complex missingness scenarios. Further, while predictive performance, fairness and stability can be seen as orthogonal, we identify trade-offs among them that arise due to the combination of missingness scenario, the choice of an imputer, and the architecture of the model trained on the data post-imputation. We make Shades-of-NULL publicly available, and hope to enable researchers to comprehensively and rigorously evaluate new missing value imputation methods on a wide range of evaluation metrics, in plausible and socially meaningful missingness scenarios.
Abstract:The error of an estimator can be decomposed into a (statistical) bias term, a variance term, and an irreducible noise term. When we do bias analysis, formally we are asking the question: "how good are the predictions?" The role of bias in the error decomposition is clear: if we trust the labels/targets, then we would want the estimator to have as low bias as possible, in order to minimize error. Fair machine learning is concerned with the question: "Are the predictions equally good for different demographic/social groups?" This has naturally led to a variety of fairness metrics that compare some measure of statistical bias on subsets corresponding to socially privileged and socially disadvantaged groups. In this paper we propose a new family of performance measures based on group-wise parity in variance. We demonstrate when group-wise statistical bias analysis gives an incomplete picture, and what group-wise variance analysis can tell us in settings that differ in the magnitude of statistical bias. We develop and release an open-source library that reconciles uncertainty quantification techniques with fairness analysis, and use it to conduct an extensive empirical analysis of our variance-based fairness measures on standard benchmarks.