Predicting Spatial Visualization Problems’ Difficulty Level from Eye-Tracking Data

Abstract

The difficulty level of learning tasks is a concern that often needs to be considered in the teaching process. Teachers usually dynamically adjust the difficulty of exercises according to the prior knowledge and abilities of students to achieve better teaching results. In e-learning, because there is no teacher involvement, it often happens that the difficulty of the tasks is beyond the ability of the students. In attempts to solve this problem, several researchers investigated the problem-solving process by using eye-tracking data. However, although most e-learning exercises use the form of filling in blanks and choosing questions, in previous works, research focused on building cognitive models from eye-tracking data collected from flexible problem forms, which may lead to impractical results. In this paper, we build models to predict the difficulty level of spatial visualization problems from eye-tracking data collected from multiple-choice questions. We use eye-tracking and machine learning to investigate (1) the difference of eye movement among questions from different difficulty levels and (2) the possibility of predicting the difficulty level of problems from eye-tracking data. Our models resulted in an average accuracy of 87.60% on eye-tracking data of questions that the classifier has seen before and an average of 72.87% on questions that the classifier has not yet seen. The results confirmed that eye movement, especially fixation duration, contains essential information on the difficulty of the questions and it is sufficient to build machine-learning-based models to predict difficulty level.

Eye Tracking for STEM Education Research: New Perspectives

“…machine learning to study (1) the differences in eye movement between different difficulty levels of the problem and (2) the possibility of predicting the difficulty level from eye-tracking data. The model generated an average accuracy of 87.60% for tracking data seen by the classifier, and 72.87% for unseen data Li et al (2020).…”