Latent space bias mitigation for predicting at-risk students

Researchers have observed the relationship between educational achievements and students' demographic characteristics in physical classroom-based learning. In the context of online education, recent studies were conducted to explore the leading factors of successful online courses. These studies also investigated how demographic features impact student achievement in the online learning environment. This motivates the use of demographic information alongside other features to predict students' academic performance. Since demographic features include protected attributes, such as gender and age, evaluating predictive models must go beyond minimizing the overall error. In this work, we analyze and investigate the use of neural networks to predict underperforming students in online courses. However, our goal is not only to enhance the accuracy but also to evaluate the fairness of the predictive models, a problem concerning the application of machine learning in education. This paper starts by analyzing the available solutions to fairness in predictive models: bias mitigation with pre-processing and in-processing methods. We show that the current evaluation is missing the case of partial awareness of protected features, which is the case when the model is aware of bias on some protected attributes but not all. The in-processing method, specifically the adversarial bias mitigation, shows that debiasing in some protected features exacerbates the bias on other protected features. This observation motivates our proposal of an alternative approach to enhance bias mitigation even in the partial awareness scenario by working with latent space. We implement the proposed solution using denoising autoencoders. The quantitative analysis used three distributions from The Open University Learning Analytics dataset (OULAD). The obtained results show that the latent space-based method offers the best solution as it maintains accuracy while mitigating the bias of the prediction models. These results indicate that in the case of partial awareness, the latent space method is considered superior to the adversarial bias mitigation approach.