In my research, I explore methods for applying machine learning and optimization techniques to improve educational technology. Topics include using existing education data to generate new hypotheses for testing, improving student experience and learning outcomes by augmenting course material with additional structure and adaptive feedback, and using rich student interaction data to understand students’ strengths and deficits and personalize the educational environment to improve outcomes.
With the advent of sophisticated machine learning techniques in recent decades came revolutions in many disciplines, including health care, business technology, and financial services. The education sector, however, remains relatively unaffected. Benefits of machine learning for personalized educational tools have been minimal, limited to simple settings such as teaching arithmetic.
In my research, I explore methods for applying machine learning and optimization techniques to improve educational technology. Topics include using existing education data to generate new hypotheses for testing, improving student experience and learning outcomes by augmenting course material with additional structure and adaptive feedback, and using rich student interaction data to understand students’ strengths and deficits and personalize the educational environment to improve outcomes.
The central setting for most of this work is the Code Seal project - a collaboration between Wolfram Language programmers and MIT subject area experts, led by Craig Carter and Kyle Keane. This team has created a tool - Code Seal - for introducing programming activities and problem sets into science courses. Code Seal is being used in physics, materials science, and math courses, both residentially at MIT and online through MITx. Code Seal is used both to expose students to programming in different contexts and to solidify their subject area understanding through visualization and computation. The extensive data collected from past courses, coupled with the opportunity to shape future iterations of courses and the development of the Code Seal tool itself has created an ideal environment for research and development of new ideas in educational technology.
My work covers a number of related topics, which I am eager to present to the educational community for thoughts and feedback.
One project looks at collections of student answers written in the Wolfram Language to generate an automated tool to provide struggling students guidance based on the trajectories of past students.
Another project builds on this work with text analysis, taking natural language from student forums to automatically generate additional structure, such as tags and graphical representations of posts. These additions aim to increase student engagement and collaboration by helping students more easily navigate to topics that are relevant to their interests.
Meanwhile, I am working on real world experiments in MIT’s introductory physics course (8.01) looking at the effects of programming and visualization assignments on student learning. In addition to experimental design and data analysis, I have used innovative approaches to stratified randomization to test groups, and I have developed new approaches to sparse and non-negative statistical techniques to aid in the generation of future hypotheses to test.
I am also early on in the development of work aiming to provide adaptive feedback to students learning to program. In this case, a much richer data source is being used, from experiments in MIT’s residential courses, including detailed time-series data on key-presses, help files, evaluations, and errors. This leads to a more novel approach using ideas from reinforcement learning, knowledge space theory, and optimization to guide student trajectories.