Knowledge Tracing Using Deep Learning Methods

Abstract

This thesis aims to address a fundamental question relating to human and machine learning: Can a machine effectively trace the dynamics of a student's knowledge states? The thesis assumes a generic definition for a student, which can be either a human learner or a machine learning model. This fundamental question lies at the heart of the Knowledge Tracing (KT) problem, which has attracted increasing interest in recent years. Addressing the KT problem is essential for a variety of applications in education, such as massive open online courses (MOOCs), intelligent tutoring systems, educational games, and learning management systems. Generally speaking, a KT model targets to achieve two main objectives. First, it needs to find effective representations of a student's knowledge states that can capture important learning concepts/skills in a learning course and their mutual relationships (e.g., prerequisite dependencies). Second, it aims to accurately trace temporal dynamics over a student's knowledge state representations which are conditioned on performance observations from their past exercise answering activities. From a machine learning perspective, a KT model is usually performed through a supervised sequence prediction task. In such tasks, a KT model sequentially predicts the probabilities of correctly answering recently observed questions conditioned on the past exercise answering history, usually considering the relationships among questions and learning concepts in the learning subject. Solving this sequence prediction task involves a number of challenges, including modeling long and short-term dependencies in an exercise answering history, finding effective ways to leverage dependency relationships among exercises, and dealing with students' forgetting behavior that may negatively impact their knowledge progress over time. To address these challenges, this thesis makes the following research contributions. First, we present a comprehensive literature review of research conducted on existing KT models and the relevant datasets and benchmarks used for model evaluation. Then, we address the challenge of modeling long-short term dependencies in exercise answering sequences by proposing a novel deep sequence learning model that can dynamically adjust its temporal attention to focus on relevant exercises for answer prediction. Following that, we consider the effect of a student's forgetting behavior during knowledge state estimate through a novel forgetting-robust model that tracks forgetting features over learning concepts during the exercise sequence modeling. We highlight the potential of KT models in evolving effective teaching strategies through a novel reinforcement learning framework, which can leverage knowledge state representations to optimize a teacher agent interactively. We comprehensively evaluate our proposed models on well-established KT datasets. Furthermore, we propose a new KT dataset that addresses existing KT datasets' limitations and make it publically available. Finally, we outline future research directions and possible application areas in the KT domain. Show more