Exploring Alternative Methods for Error Attribution in Learning Curves Analysis in Intelligent Tutoring Systems

  • Authors:

  • Adaeze Nwaigwe;Kenneth R. Koedinger;Kurt Vanlehn;Robert Hausmann;Anders Weinstein

  • Affiliations:

  • Human-Computer Interaction Institute, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA, anwaigwe@cs.cmu.edu, koedinger@cmu.edu;Human-Computer Interaction Institute, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA, anwaigwe@cs.cmu.edu, koedinger@cmu.edu;LRDC, University of Pittsburgh, PA, USA, [vanlehn, bobhaus, andersw]@pitt.edu;LRDC, University of Pittsburgh, PA, USA, [vanlehn, bobhaus, andersw]@pitt.edu;LRDC, University of Pittsburgh, PA, USA, [vanlehn, bobhaus, andersw]@pitt.edu

  • Venue:
  • Proceedings of the 2007 conference on Artificial Intelligence in Education: Building Technology Rich Learning Contexts That Work
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

It is becoming a standard technique to use learning curves as part of evaluation of intelligent tutoring systems [1,2,3], but such learning curves require a method for attributing errors. That is, the method must determine for each error a student makes what “knowledge component” in the student model is to blame. To this point, alternative methods for error attribution have not been systematically investigated. We implemented four alternative methods for error attribution --two temporal heuristics and two location-temporal heuristics. We employed two evaluation standards --a statistical standard for measuring model fit and parsimony and the Kappa technique for measuring inter-observer reliability. We looked to see which method better met the “learning-curve standard” that is, led to better prediction of students' changes in error rate over time. Second, we asked if the codes generated by the methods better met the “human-match standard”, that is, were they like error attributions made by human coders. Both evaluation standards led to better results for the location-temporal heuristic methods than the temporal heuristic methods. Interestingly, we found that two of the methods proposed were better at error attribution, according to the learning curve standard, than the original cognitive model of the intelligent tutoring system. Overall, these results suggest that the heuristics proposed and implemented in this paper can generally aid learning curve analysis and perhaps, more generally, the design of student models