Evaluation of the instance weighting strategy for transfer learning of educational predictive models

Mariia Luzan, Christopher Brooks

Track: Responsible AI for Education (Day 2)

Paper Length: long-paper (6 pages + references)

Keywords: Transfer learning; Higher education; Student dropout prediction; Fairness

Abstract:

This work contributes to our understanding of how transfer learning can be used to improve
educational predictive models across higher institution units. Specifically, we provide an
empirical evaluation of the instance weighting strategy for transfer learning, whereby a
model created from a source institution is modified based on the distribution characteristics
of the target institution. In this work we demonstrated that this increases overall model
goodness-of-fit, increases the goodness-of-fit for each demographic group considered, and
reduces disparity between demographic groups when we consider a simulated institutional
intervention that can only be deployed to 10% of the student body.

Submission Number: 44