Algorithmic Bias in a Student Success Prediction Models: Two Case Studies

Machine learning algorithms are increasingly being used in today’s society. However, growth in these algorithms means growth in algorithmic bias, and it is imperative that we work to understand the bias that may result. One area of study in which these algorithms are widely used is in educational institutions. The algorithms are often used to predict student success or retention. In our research, we aim to uncover the biases that may result from building and using a machine learning student success models. To do so, we used two publicly available student datasets from educational settings (one from a MOOC and another one from secondary education in Portugal) and built models of our own. We then compared the accuracy and the fairness of each model type to observe which models performed best on each subcategory of students. Among the models we built, we found that while it is easy to use accuracy to evaluate models and find the most accurate ones, the most accurate predictive model overall for a dataset may not perform fairly in predicting student success for all subcategories of students. To better tune models for fairness, we found that it is possible to tune models that also take fairness into consideration, and these models could perform more fairly on almost all subcategories of students, but it slightly took away from the accuracy of the algorithm. Our results demonstrate the importance of creating and tuning several model types in order to choose a balanced model that balances accuracy and fairness.