•   When: Monday, April 27, 2020 from 10:00 AM to 12:00 PM
•   Speakers: Qian Hu
•   Location: Webex
•   Export to iCal

Higher educational institutions face major challenges including timely graduation and retention of enrolled students. The National Center for Education Statistics (NCES) reports that the six-year graduation rate for first-time and full-time undergraduates is around 60%; the retention rate among first-time and full-time degree-seeking students is around 80%. These alarming statistics require higher educational institutions to take actions to improve their effectiveness and efficiency at educating students. Educational data mining technologies for academic trajectory and degree planning, course recommender systems, early warning and advising seek to improve student success. The foundation of these systems is student modeling and forecasting. 
 
However, developing appropriate and accurate predictive models for modeling students is a non-trivial problem due to several challenges. The first challenge is that a student’s learning is influenced by many factors such as motivation, affect and identify. It is further compounded by the fact that learning is a reflection of cognition which is not a simple process. Students can also choose to take courses in different sequences at different pace. The second challenge is that degree programs exhibit complex knowledge dependence between courses. When it comes to decision making, machine learning has its shortcomings in terms of predictive reliability and interpretability. A reliable model is able to express its prediction confidence so that human decision-makers can know when the predictions are trustworthy. Interpretable models can provide explanations for predictions and decision-makers can use the explanations to guide their decisions. Recently, several concerns have emerged about the fairness of using machine learning models. A biased predictive model may negatively influence subgroups of the larger population. For example, in the educational setting, models unfairly predicting a particular group of students to be at a higher risk of failure can discourage them from pursuing their degree pathway.
 
In this thesis, we develop novel and accurate machine learning models for student modeling and forecasting. Specifically, we develop sequential and graph machine learning models to model students' learning processes and predict their academic performance. Towards informed decision making, we develop Bayesian deep learning models to quantify uncertainty. We also propose a metric-free individual fairness formalization and develop two fair machine learning models using neural classifiers and gradient contextual bandits for mitigating unfairness in these predictive models.