Ensemble Learning With Probabilistic Models of Student Knowledge in MITutor

• In Event: Technical Issues in Data Mining and Learning Analytics

Abstract

The aim of this study is to examine an ensemble approach to model skill acquisition in MITutor, an intelligent tutoring system that engages students in dialogue to practice motivational interviewing with virtual agents. The proposed ensemble leverages variants of simple probabilistic models that serve as outputs to a meta-learner in addition to domain-related variables. We trained the ensemble model from 1,961 responses made by 66 students to virtual agents and compared its performance to a single, best performing probabilistic model following a student-level cross-validation procedure. The ensemble outperformed the single classifier with 80.68% accuracy. We discuss the implications for adaptive task sequencing to optimize the amount of practice for skill acquisition.

Authors

• Kent Ellsworth

• Eric G. Poitras, Dalhousie University

• Zac Imel, The University of Utah

• Robert Zheng, University of Utah

• Derek Caperton

• Grin Lord

• Jake Van Epps, The University of Utah

• Michael Tanana, The University of Utah

• David Atkins, University of Washington