Ranking feature sets for emotion models used in classroom based intelligent tutoring systems

摘要: Recent progress has been made by using sensors with Intelligent Tutoring Systems in classrooms order to predict the affective state of students users If tutors are able interpret sensor data new based on past experience, rather than having be individually trained, then this will enable tutor developers evaluate various methods adapting each student's consistent predictions In past, our classifiers have predicted student emotions an accuracy between 78% and 87% However, it is still unclear which best, educational technology community needs know develop better baseline classifiers, e.g ones that use only frequency emotional occurrence This paper suggests a method clarify classifier ranking for purpose models The begins careful collection training testing set, from separate population, concludes non-parametric trained set We illustrate collected Fall 2008 tested Spring 2009 Our results show some states significantly model; validation analysis showed but not all rankings generalize settings Overall, though there benefit gained simple linear more advanced or features may needed classification performance.