Confidence binary classification using eye gaze information based on temporal self-evaluations

Abstract

In cooperative learning with a communication robot, it is important to let the robot understand a person’s mental state, such as emotion and confidence. We aim to estimate the confidence level from eye gaze information using machine learning to develop a system that provides hints based on the learner’s mental state. In this paper, we introduced temporal self-evaluation by the participants themselves using a touchscreen slider. Based on the data collected from participants’ responses to 30 questions, we used Support Vector Machine (SVM) to classify “Unconfident” and “Confident,” and found that the classifica-(breakpoint)tion accuracy was 94.4% on average for one specific participant, but 66.8% when one participant’s data was tested based on other participants’ data. While we achieved the result that classification by learning the data of a specific individual is suitable, we also found that the score decreased because of individual differences such as data bias, and therefore, it is an issue to revise the analysis method.