Personalizing education with machine learning

The field of education has the potential to be transformed by the internet and intelligent computer systems. Evidence for the first stage of this transformation is abundant, from the Stanford online AI and Machine Learning courses to web sites such as Kahn Academy that offer on line lessons and drills. However, the delivery of instruction via web-connected devices is merely a precondition for what may become an even more fundamental transformation: the personalization of education.

In traditional classroom settings, teachers must divide their attention and time among many students and hence have limited ability to observe and customize instruction to individuals. Even in one-on-one tutoring sessions, teachers rely on intuition and experience to choose the material and stye of instruction that they believe would provide the greatest benefit given the student's current state of understanding.


In order both to assist human teachers in traditional classroom environments and to improve automated tutoring systems to match the capabilities of expert human tutors, one would like to develop formal approaches that can:

* exploit subtle aspects of a student's behavior---such as facial expressions, fixation sequences, response latencies, and errors---to make explicit inferences about the student's latent state of knowledge and understanding;

* leverage the latent state to design teaching policies and methodologies that will optimize the student's knowledge acquisition, retention, and understanding; and

* personalize instruction by providing material and interaction suited to the capabilities and preferences of the student.

Machine learning provides a rich set of tools, extending classical psychometric approaches, for data-driven latent state inference, policy optimization, and personalization. Years ago, it would have been difficult to obtain enough data for a machine learning approach. However, online interactions with students have become commonplace, and these interactions yield a wealth of data. The data to be mined go beyond what is typed: Cameras and microphones are ubiquitous on portable devices, allowing for the exploitation of subtle video and audio cues. Because web-based instruction offers data from a potentially vast collection of diverse learners, the population of learners should serve useful in drawing inferences about individual learners.

Mining the vast datasets on teaching and learning that emerge over the coming years may both yield important insights into effective teaching strategies and also deliver practical tools to assist both human and automated teachers.

The goal of this workshop is to bring together researchers in machine learning, data mining, and computational statistics with researchers in education, psychometrics, intelligent tutoring systems, and designers of web-based instructional software. Although a relatively young journal and conference on educational data mining has been established (educationaldatamining.org), the field hasn't had as much contact with machine learning theoreticians as one would like.