The Accuracy Paradox of Algorithmic Classification

Abstract

In recent years, algorithmic classification based on machine learning techniques has been increasingly permeating our lives. With their increased ubiquity, negative social consequences have come to light. Among these consequences are ´unfair´ algorithms. This resulted in a large body of research tackling ´fairness´ of algorithms and related issues. Algorithms are frequently considered as unfair if they show diverging accuracies for different groups, with a particular focus on vulnerable groups, indicating a correlation between prediction and information about group membership. In this paper I argue that, while this research contributes valuable insights, much of the research focuses a quantitative understanding of fairness which creates a very narrow focus. My argument builds on four pillars. First, much of the research on 'fairness' focuses on accuracy as basis for ´fairness´. Even though ´fairness´ can reduce the overall accuracy, this is seen as a limitation, implicitly aiming for high accuracy. Second, this focus is in line with other debates about algorithmic classification that focus on quantiative performance measures. Third, close attention on accuracy may be a pragmatic and well-intended stance for practicioners but can distract from problematizing the ´bigger picture´. Fourth, I argue that any classification produces a marginalized group, namely those that are misclassified. This marginalization increases with the classifier´s accuracy, and in tandem the ability of the affected to challenge the classification is diminished. Combined, this leads to the situation that a focus on fairness and accuracy may weaken the position and agency of those being misclassified, paradoxically contradicting the promissory narrative of ´fixing´ algorithms through optimizing fairness and accuracy.