Thursday, September 29, 2022, 4pm

Intelligent tutoring systems (ITSs) are an educational technology which provides millions of learners worldwide with access to learning materials and personalized instruction. At their core, ITSs depend on their ability to assess the student's evolving degree of proficiency to adapt the learning process by selecting effective instructional materials at each point in time. As we do not have a way to directly observe the student's latent knowledge state, machine learning algorithms are used to make inferences about the student's knowledge state based on sequence log data that describes the student's interactions with the ITS. In this thesis we study the question of how the large-scale data that is available in today's learning systems can be used to give those systems better methods for student assessments and for choosing the right teaching action for the individual student.

In the first part we focus on problems related to student assessments via student performance models (SPMs). We explore how to enhance SPM accuracy by leveraging alternative types of student data that go beyond conventional question-answering logs and by learning beneficial question embeddings via pre-training. ITSs face cold-start problems when new content is added and when new users enter the system. We introduce transfer learning techniques that use log data from existing courses to mitigate the new content cold-start problem for new courses. We address the new user cold-start problem by exploring algorithms for early success predictions that estimate a student's expected performance and time requirements before starting a new topic. The second part of this thesis asks how to bridge the gap between estimating a student's state to adapting the learning process to the student by selecting effective teaching actions. Based on the use case of a real-world online ITS frequented by hundreds of thousands of K-12 students we explore ways in which bandit and reinforcement learning techniques can leverage student state estimates to learn effective teaching policies.

Thesis Committee:
Tom Mitchell (Chair)
Barnabás Póczos
Vincent Aleven
Min Chi (North Carolina State University)

In Person and Zoom Participation. See announcement.

Event Type: Thesis Proposals
Room Number: In Person and Virtual - ET
Building: Gates Hillman 7101 and Zoom
Speaker's Name: ROBIN SCHMUCKER
Speaker Website: rschmucker.github.io
Speaker's Professional Title: Ph.D. Student, Machine Learning Department, Carnegie Mellon University
Talk Title: Sequence-Modeling for Assessments and Interventions in Intelligent Tutoring Systems
For More Information: stidle@andrew.cmu.edu
Affiliations: Machine Learning Department (MLD)
Organization(s): SCS