Should Course-based At-risk Predication Models Include Protected Features?
Moohanad Jawthari, Veronika Stoffa · 2022
Nowadays, online learning is preferable as it allows learners to learn anytime and anywhere, especially in pandemic crisis. However, this type of learning suffers from high dropout and low completion rates. Identifying at-risk students can improve students’ success and institutes effectiveness. Therefore, this paper proposes an at-risk predictive model by exploiting students behavioral and assessment data using data from a large distance-based university, Open University in UK. Unlike the tradition way of training and testing a model using one course, the proposed model is trained based on three historical course data and tested on the last offering. The prediction models show high performance values like 0.91 in accuracy. At same time, the paper examines if including protected feature like gender make prediction models discriminates against underrepresented student groups and aggravate present inequalities. To achieve this, different fairness metrics are used to compare models predictions with and without demographic characteristics. We find including of demographic characteristics does affect prediction performance of the models. In terms of fairness, there is no bias towards male and female students’ groups, but models can not be called fair for other features as there is bias in at least one metric towards other attributes.