Measuring algorithmic bias in machine learning has historically focused on statistical inequalities pertaining to specific groups. However, the most common metrics (i.e., those focused on individual- or group-conditioned error rates) are not currently well-suited to educational settings because they assume that each individual observation is independent from the others. This is not statistically appropriate when studying certain common educational outcomes, because such metrics cannot account for the relationship between students in classrooms or multiple observations per student across an academic year. In this paper, we present novel adaptations of algorithmic bias measurements for regression for both independent and nested data structures. Using hierarchical linear models, we rigorously measure algorithmic bias in a machine learning model of the relationship between student engagement in an intelligent tutoring system and year-end standardized test scores. We conclude that classroom-level influences had a small but significant effect on models. Examining significance with hierarchical linear models helps determine which inequalities in educational settings might be explained by small sample sizes rather than systematic differences.

Recent research has made strides toward fair machine learning. Relatively few datasets, however, are commonly examined to evaluate these fairness-aware algorithms, and even fewer in education domains, which can lead to a narrow focus on particular types of fairness issues. In this paper, we describe a novel dataset modification method that utilizes a genetic algorithm to induce many types of unfairness into datasets. Additionally, our method can generate an unfairness benchmark dataset from scratch (thus avoiding data collection in situations that might exploit marginalized populations), or modify an existing dataset used as a reference point. Our method can increase the unfairness by 156.3% on average across datasets and unfairness definitions while preserving AUC scores for models trained on the original dataset (just 0.3% change, on average). We investigate the generalization of our method across educational datasets with different characteristics and evaluate three common unfairness mitigation algorithms. The results show that our method can generate datasets with different types of unfairness, large and small datasets, different types of features, and which affect models trained with different classifiers. Datasets generated with this method can be used for benchmarking and testing for future research on the measurement and mitigation of algorithmic unfairness.

Abstract Automated, data‐driven decision making is increasingly common in a variety of application domains. In educational software, for example, machine learning has been applied to tasks like selecting the next exercise for students to complete. Machine learning methods, however, are not always equally effective for all groups of students. Current approaches to designing fair algorithms tend to focus on statistical measures concerning a small subset of legally protected categories like race or gender. Focusing solely on legally protected categories, however, can limit our understanding of bias and unfairness by ignoring the complexities of identity. We propose an alternative approach to categorization, grounded in sociological techniques of measuring identity. By soliciting survey data and interviews from the population being studied, we can build context‐specific categories from the bottom up. The emergent categories can then be combined with extant algorithmic fairness strategies to discover which identity groups are not well‐served, and thus where algorithms should be improved or avoided altogether. We focus on educational applications but present arguments that this approach should be adopted more broadly for issues of algorithmic fairness across a variety of applications.

In recent years, machine learning has become more common in everyday applications. Consequently, numerous studies have explored issues of unfairness against specific groups or individuals in the context of these applications. Much of the previous work on unfairness in machine learning has focused on the fairness of outcomes rather than process. We propose a feature selection method inspired by fair process (procedural fairness) in addition to fair outcome. Specifically, we introduce the notion of unfairness weight, which indicates how heavily to weight unfairness versus accuracy when measuring the marginal benefit of adding a new feature to a model. Our goal is to maintain accuracy while reducing unfairness, as defined by six common statistical definitions. We show that this approach demonstrably decreases unfairness as the unfairness weight is increased, for most combinations of metrics and classifiers used. A small subset of all the combinations of datasets (4), unfairness metrics (6), and classifiers (3), however, demonstrated relatively low unfairness initially. For these specific combinations, neither unfairness nor accuracy were affected as unfairness weight changed, demonstrating that this method does not reduce accuracy unless there is also an equivalent decrease in unfairness. We also show that this approach selects unfair features and sensitive features for the model less frequently as the unfairness weight increases. As such, this procedure is an effective approach to constructing classifiers that both reduce unfairness and are less likely to include unfair features in the modeling process.