Machine Learning Techniques for Personalized Learning
Doctor of Philosophy
Recent developments in personalized learning, powered by recent advances in machine learning and big data, have the potential to revamp the “one-size-fits-all” approach in today’s education by delivering a fully personalized learning experience for each student. The key behind these developments is to create a personalized learning system (PLS), which can automatically deliver analytics and feedback on the students’ progress and recommend learning actions for the students to take. A PLS presents a scalable approach to personalized learning by analyzing the data students generate while interacting with learning resources (i.e., textbook sections, lecture videos, assessment questions, etc). Such an approach relies on only minimal human effort and has the ability to scale to applications with millions of students, thousands of learning resources, and hundreds of domains. In this thesis, we develop a series of machine learning techniques for personalized learning, building on our previous work on sparse factor analysis (SPARFA) for learning and content analytics. To begin with, we develop a new, nonlinear latent variable model that we call the dealbreaker model, in which a student’s success probability is determined by their weakest concept mastery. We develop efficient parameter inference algorithms for this model using novel methods for nonconvex optimization. We demonstrate that the dealbreaker model excels at prediction and the parameters learned are interpretable: they provide key insights into which concepts are critical (i.e., the “dealbreakers”) to answering a question correctly. We also apply the dealbreaker model to a movie rating dataset, illustrating its broad applicability to applications other than education. Then, we propose SPARFA-Trace, a new framework for time-varying learning and content analytics. We develop a novel message passing-based, blind, approximate Kalman filtering and smoothing algorithm for SPARFA that jointly traces student concept knowledge evolution over time, analyzes student concept knowledge state transitions (induced by studying learning resources, such as textbook sections, lecture videos, etc., or the forgetting effect), and estimates the content organization and difficulty of the questions in assessments. These quantities are estimated solely from binary-valued (correct/incorrect) graded student response data and the specific actions each student performs (e.g., answering a question or studying a learning resource) at each time instant. Additionally, we study the problem of automatic grading and feedback generation for the kinds of open response mathematical questions that figure prominently in STEM (science, technology, engineering, and mathematics) courses. Our data-driven framework for mathematical language processing (MLP) leverages solution data from a large number of students to evaluate the correctness of their solutions, assign partialcredit scores, and provide feedback to each student on the likely locations of any errors. MLP takes inspiration from the success of natural language processing for text data and comprises three main steps. First, we convert each solution to an open response mathematical question into a series of numerical features. Second, we cluster the features from several solutions to uncover the structures of correct, partially correct, and incorrect solutions. We develop two different clustering approaches, one that leverages generic clustering algorithms and one based on Bayesian nonparametrics. Third, we automatically grade the remaining (potentially large number of) solutions based on their assigned cluster and one instructor-provided grade per cluster. As a bonus, we can track the cluster assignment of each step of a multistep solution and determine when it departs from a cluster of correct solutions, which enables us to indicate the likely locations of errors to students. Furthermore, we study the problem of selecting the best personalized learning action that each student should take next given their learning history; actions could include reading a textbook section, watching a lecture video, interacting with a simulation or lab, solving a practice question, and so on. We first estimate each student’s knowledge profile from their binary-valued graded responses to questions in their previous assessments. We then employ these knowledge profiles as contexts in the contextual (multi-armed) bandits framework to learn a policy that selects the personalized learning actions that maximize each student’s immediate success, i.e., their performance on their next assessment. We develop three algorithms for personalized learning action selection. While one is mainly of theoretical interest, we experimentally validate the other two using real-world educational datasets. Our proposed set of models and algorithms comprise the basic and most essential components of a PLS, i.e., learning analytics, content analytics, grading and feedback, and scheduling.