COMMITTEE CHAIR: Dr. Lin Li
TITLE: INTEGRATING SEMI-SUPERVISED LEARNING AND ENSEMBLE DEEP LEARNING FOR LOW-RESOURCE DEEP KNOWLEDGE TRACING
ABSTRACT: As learning platforms scale, two obstacles keep Deep Knowledge Tracing (DKT) from thriving in practice: scarce labeled interactions and opaque predictions. We introduce a semi-supervised, ensemble DKT pipeline that trains diverse architectures, including Knowledge Proficiency Tracing (KPT), Exercise-Correlated KPT (EKPT), and Dynamic Key-Value Memory Networks (DKVMN) on a mixture of limited labeled and abundant unlabeled learner interactions, then aggregates predictions via majority voting to stabilize learning and reduce variance. Evaluated on benchmark datasets such as ASSISTments, the combined approach yields consistent gains in AUC, accuracy, and precision over strong supervised baselines while revealing influential interactions, surfacing spurious correlations and temporal leakage risks, and guiding targeted interventions. This contribution provides data-efficient pipelines that enhance predictive performance, foster educator trust, and support personalized learning on a scale.
Room Location: S.R. Collins Building, Room 111L