Fair and Actionable Early-Warning Analytics for Student Dropout in Engineering Programs
Fair and Actionable Early-Warning Analytics for Student Dropout in Engineering Programs
Description
Details
Context and Problem Statement
Early-warning systems often use factors that institutions cannot change, such as socioeconomic background, creating potentially unfair and inactionable alerts. Restricting features to modifiable behaviors may reduce apparent accuracy but increase practical value.
Research Question
What predictive-performance cost results from restricting a dropout-warning model to actionable variables, and does this restriction improve fairness across student subgroups?
Proposed Approach
Use knowledge tracing and survival analysis, compare unrestricted and actionable-feature models, audit subgroup fairness, and analyze proposed interventions counterfactually.
Expected Contribution
A quantitative characterization of the trade-off among prediction, fairness, and actionability.
Expected Prototype
A dashboard for academic coordinators showing at-risk groups, actionable factors, and support recommendations without stigmatizing individual scores.
Datasets
OULAD, EdNet, and anonymized institutional data under ethical approval.
Challenges
Stigmatization, self-fulfilling effects, privacy, small cohorts, and longitudinal validation.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Machine Learning and Survival Analysis
- Statistics and Fairness Auditing
- Educational Sciences
- Ethics and Data Protection
Datasets
- OULAD (Open University Learning Analytics Dataset)
- EdNet
- Anonymized institutional data under formal agreement