A Bayesian Multi-scale Personalized Learning Approach
Abstract
This 3-year project is concerned with finding a scalable way of realizing the potential benefits of personalized learning by increasing the number and diversity of learning resources -- such as explanations of key concepts, worked examples, and hints -- that such platforms can serve to their students to help them learn. In particular, we wish to create a crowdsourcing-based learningcommunity whereby both teachers and students can author and share novel learning resources (LRs) with other community members. By harnessing key interactions between features of the resources and the state of the learner, each student can receive an optimal set of resources to maximize their expected learning gains. In order to identify quickly which learning resources aremore effective, a multi-timescale Bayesian optimization approach will be used that combines subjective rating information (short time-scale) with more distal, but objective, learning outcome information (longer time-scale). Crowdsourcing new learning resources represents a significant departure from the classical approach to personalized learning: instead of tuning parameters for afixed set of possible teaching actions (give hint, increase difficulty, etc.), the proposed approach is to grow the set of teaching actions itself. The goal of our project is inspired in part by the success of volunteer-based educational communities such as Wikipedia and StackOverflow. From a technical perspective, the proposed research will consist of innovations incrowdsourcing, multimodal machine learning, and Bayesian optimization. The crowdsourcing research literature contains two main branches: a human-computer interaction focus that investigates how to create user interfaces and workflows that induce many workers to contribute high-quality content, as well as a machine learning focus that explores how the more reliablecontributors can be identified automatically and how multiple contributions can be combined to maximize some objective function. Our research will span both branches. From an applied multimodal machine learning perspective, one of the goals of the proposed research is to automatically extract key features (e.g., from clickstream logs, images of students??? math solutions, and textual open-responses) -- such as the affective state of the user, and linguisticcomplexity of an authored explanation or hint -- that are useful for deciding which resource to give to which student. From a Bayesian optimization perspective, we will go beyond prior work on applying Bayesian optimization methods to human learning by harnessing not just the main effects of, but also harnessing interactions between, the key features of the resources and states of the learner.Research and development will take place using the ASSISTments intelligent tutoring system, which serves 50,000 learners from both middle- and high school. We will use the co-design methodology, in which we work closely with 4 high school teachers to identify the best ways to implement the proposed platform. The research project will build on prior work and software that was conducted in an ASSISTments-based pilot project called "PeerASSIST" inwhich we have started testing out user interface designs. The research team has the necessary interdisciplinary expertise in learning science, education, machine learning, and crowdsourcing to tackle this project.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 04, 2018
- Source ID
- N000141812768
Entities
People
- Neil Heffernan
Organizations
- Office of Naval Research
- United States Navy
- Worcester Polytechnic Institute