Search for: All records

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.

 

Recent research seeks to develop more comprehensive learner models for adaptive learning software. For example, models of reading comprehension built using data from students’ use of adaptive instructional software for mathematics have recently been developed. These models aim to deliver experiences that consider factors related to learning beyond performance in the target domain for instruction. We investigate the extent to which generalization is possible for a recently developed predictive model that seeks to infer students’ reading comprehension ability (as measured by end-of-year standardized test scores) using an introductory learning experience in Carnegie Learning’s MATHia intelligent tutoring system for mathematics. Building on a model learned on data from middle school students in a single school district in a mid-western U.S. state, using that state’s end-of-year English Language Arts (ELA) standardized test score as an outcome, we consider data from a school district in a south-eastern U.S. state as well as that state’s end-of-year ELA standardized test outcome. Generalization is explored by considering prediction performance when training and testing models on data from each of the individual school districts (and for their respective state’s test outcomes) as well as pooling data from both districts together. We conclude with discussion of investigations of some algorithmic fairness characteristics of the learned models. The results suggest that a model trained on data from the smaller of the two school districts considered may achieve greater fairness in its predictions over models trained on data from the other district or both districts, despite broad, overall similarities in some demographic characteristics of the two school districts. This raises interesting questions for future research on generalizing these kinds of models as well as on ensuring algorithmic fairness of resulting models for use in real-world adaptive systems for learning. 

Research into "gaming the system" behavior in intelligent tutoring systems (ITS) has been around for almost two decades, and detection has been developed for many ITSs. Machine learning models can detect this behavior in both real-time and in historical data. However, intelligent tutoring system designs often change over time, in terms of the design of the student interface, assessment models, and data collection log schemas. Can gaming detectors still be trusted, a decade or more after they are developed? In this research, we evaluate the robustness/degradation of gaming detectors when trained on old data logs and evaluated on current data logs. We demonstrate that some machine learning models developed using past data are still able to predict gaming behavior from student data collected 16 years later, but that there is considerable variance in how well different algorithms perform over time. We demonstrate that a classic decision tree algorithm maintained its performance while more contemporary algorithms struggled to transfer to new data, even though they exhibited better performance on both new and old data alone. Examining the feature importances provides some explanation for the differences in performance between models, and offers some insight into how we might safeguard against detector rot over time.