Educational Data Mining
W. L. Miller, R. Baker, M. Labrum, K. Petsche, Y.-H. Liu, and A. Z. Wagner. The Fifth International Learning Analytics & Knowledge Conference (LAK 15) (2015).
In collaboration with Columbia University Teachers College professor Ryan Baker, we build automated, sensor-free detectors of “proactive remediation” (on-the-spot teacher intervention with students based upon their performance) by Reasoning Mind teachers. Previous work has focused on student behaviors; here, we are able to infer teacher behavior based upon student actions. In particular, we can distinguish them from student off-task behavior or idle time. This and the following EDM papers use a variety of machine learning techniques to build models of student and teacher behavior and affect, based wholly on the student’s activities within the system.
W. L. Miller, R. Baker, M. J. Labrum, K. Petsche and A. Z. Wagner. KDD Workshop on Data Mining for Educational Assessment and Feedback (ASSESS 2014) (2014).
We construct an automated, sensor-free detector of student boredom within the Readoning Mind system. To demonstrate the use of this detector, we apply it to data from 70,000 students across the entire school year. Analysis of the results reveals a systematic, bimodal distribution in student boredom; the Reasoning Mind curriculum is roughly evenly split between high- and low-boredom lessons.
Unifying Computer-Based Assessment Across Conceptual Instruction, Problem-Solving, and Digital Games
W. L. Miller, R. S. Baker, and L. M. Rossi. Technology, Knowledge and Learning 19, 165 (2014). (preprint)
As students work through online learning systems such as Reasoning Mind, they often are not confined to working within a single educational activity; instead, they work through various different activities such as conceptual instruction, problem-solving items, and fluency-building games. In this paper, we apply Bayesian Knowledge Tracing (BKT), which is a method of estimating a student’s knowlege of different topics (known as knowledge components) in an educational system. In particular, we explore different ways of modeling student knowledge across the different types of activies they encounter, and find that integrating different activities does not improve predictive performance.
As part of this work, I adapted Baker et al.’s BKT Brute Force algorithm to simulated annealing for parameter fitting, resulting in a speed-up of roughly fifty times. A Java implementation of this method can be seen on GitHub here.
Soft Matter Physics
W. L. Miller, B. Bozorgui, K. Klymko, and A. Cacciuto. Journal of Chemical Physics 135, 244902 (2011). (preprint)
W. L. Miller and A. Cacciuto. Soft Matter 7, 7552 (2011). (preprint)
W. L. Miller and A. Cacciuto. Journal of Chemical Physics 133, 234903 (2010). (preprint)
W. L. Miller and A. Cacciuto. Journal of Chemical Physics 133, 234108 (2010). (preprint)
W. L. Miller, B. Bozorgui, and A. Cacciuto. Journal of Chemical Physics 132, 134901 (2010). (preprint)
B. Bozorgui, M. Sen, W. L. Miller, J. C. Pàmies, and A. Cacciuto. Journal of Chemical Physics 132, 014901 (2010). (preprint)
W. L. Miller and A. Cacciuto. Physical Review E 80, 021404 (2009). (preprint)
An overview of this work can be found in my undergraduate thesis, which can be downloaded here (PDF link).
R. E. Palacios, W.-S. Chang, J. K. Grey, W. L. Miller, C.-Y. Liu, G. Henkelman, D. Zepeda, J. Ferraris, and P. F. Barbara. Journal of Physical Chemistry B 113, 14619 (2009).
S.-J. Park, S. Link, W. L. Miller, A. Gesquiere, and P. F. Barbara. Chemical Physics 341, 169 (2007).
J. K. Grey, D. Y. Kim, B. C. Norris, W. L. Miller, and P. F. Barbara. Journal of Physical Chemistry B 110, 25568 (2006).
Effect of temperature and chain length on the bimodal emission properties of single polyfluorene copolymer molecules
J. K. Grey, D. Y. Kim, C. L. Donley, W. L. Miller, J. S. Kim, C. Silva, R. H. Friend, and P. F. Barbara. Journal of Physical Chemistry B 110, 18898 (2006).