Data-driven Methods for Course Selection and Sequencing

University of Minnesota Ph.D. dissertation.May 2019. Major: Computer Science. Advisor: George Karypis. 1 computer file (PDF); xiii, 115 pages.Learning analytics in higher education is an emerging research field that combines data mining, machine learning, statistics, and education on learning-related data, in order to develop methods that can improve the learning environment for learners and allow educators and administrators to be more effective. The vast amount of data available about students' interactions and their performance in classrooms has motivated researchers to analyze this data in order to gain insights about the learning environment for the ultimate goal of improving undergraduate education and student retention rates. In this thesis, we focus on the problem of course selection and sequencing, where we would like to help students make informed decisions about which courses to register for in their following terms. By analyzing the historical enrollment and grades data, this thesis studies the two main problems of course selection and sequencing, namely grade prediction and course recommendation. In addition, it analyzes the relationship between degree planning in terms of course timing and ordering and the students' GPA and time to degree. First, we focus on predicting the grades that students will obtain on future courses so that they can make informed decisions about which courses to register for in their following terms. We model the grade prediction problem as cumulative knowledge-based linear regression models that learn the courses' required and provided knowledge components and use them to estimate a student's knowledge state at each term and predict the grades that he/she can obtain on future courses. Second, we focus on improving the knowledge-based regression models we previously developed by modeling the complex interactions among prior courses using non-linear and neural attentive models, in order to have more accurate estimation of a student's knowledge state. In addition, we model the interactions between a target course, which we would like to predict its grade, and the other courses taken concurrently with it. We hypothesize that concurrently-taken courses can affect a student's performance in a target course, and thus modeling their interactions with that course should lead to better predictions. Third, we focus on analyzing the degree plans of students to gain more insights about how course timing and sequencing relate to their GPAs and time to degree. Toward this end, we define several course timing and course sequencing metrics and compare different sub-groups of students who have achieved high vs low GPA as well as sub-groups of students who have graduated on time vs over time. Fourth, we focus on improving course recommendation by recommending to each student a set of courses which he/she is prepared for and expected to perform well in. We model this problem as a grade-aware course recommendation problem, where we propose two different approaches. The first approach ranks the courses by using an objective function that differentiates between courses that are expected to increase or decrease a student's GPA. The second approach combines the grades predicted by grade prediction methods with the rankings produced by course recommendation methods to improve the final course rankings. To obtain the course rankings in both approaches, we adapted two widely-used representation learning techniques to learn the optimal temporal ordering between courses. In summary, this thesis addresses two closely related problems by: (1) developing cumulative knowledge-based regression models for grade prediction; % (2) developing context-aware non-linear and neural attentive knowledge-based models for grade prediction; % (3) analyzing degree planning and how the time when students take courses and how they sequence them relate to their GPAs and time to degree; and % (4) developing novel approaches for grade-aware course recommendation.