One Decision Tree is Enough to Make Customization
Hao Wan and Joseph Beck
The ability to customize instruction to individuals is a great potential for adaptive educational software. Unfortunately, beyond mastery learning and learner control, there has not been much work with adapting instruction to individuals. This paper provides an approach to determine what type of learner does best with a different intervention. We focused on constructing a decision tree that discriminated difference between tutoring interventions, and thus to make customization for each student. This paper also focuses more broadly than a particular intervention, and analyzes trends across 22 experimental interventions with a computer-based tutor for mathematics. We evaluated our model on simulated and on real data. In the simulated data set, it outperformed other methods and the constructed models captured a pre-defined customization structure. With the real data, the customized learning approach achieved stronger learning gains than simply picking the best overall teaching option. Surprisingly, it was difficult to outperform a decision tree that simply used how quickly students tended to learn a skill. That is, more features and more complex models did not result in a more effective system.