Educational Data Mining
We have recently turned our attention to analyses of students' behaviors as they interact with our teachable agents. Such analyses shed light on students' choices of behaviors that influence learning, and the strategies they bring to the task. Preliminary analyses have shown that the quality of students' concept maps are paralleled by patterns in their behaviors.
In our analyses, we have refined a methodology for exploring students' strategies using hidden Markov models (HMMs) to describe patterns of student behavior. Student actions are extracted from extensive log files and mined using these statistical learning methods. The resulting models provide an aggregrate summary of students' interactions with the system. Much like factor analysis, it is up to the researcher to give meaning to the derived states.
Our interpretations of the abstract HMMs have given rise to several patterns that we feel representant meaningful strategies used by students: basic map building, map probing, and map tracing. These aggregates describe increasing levels of sophistication and metacognition as students move from adding concepts to their maps (map building) to examining the reasoning and explanations implied by their map structures (map tracing).
The image below depicts patterns of these activities across three conditions in a recent study (ICS = no teaching, LBT = teaching with content feedback, and SRL = teaching with metacognitive feedback).
