Estimating Self-Confidence in Video-Based Learning Using Eye-Tracking and Deep Neural Networks

Abstract

Self-confidence is a crucial trait that significantly influences performance across various life domains, leading to positive outcomes by enabling quick decision-making and prompt action. Estimating self-confidence in video-based learning is essential as it provides personalized feedback, thereby enhancing learners' experiences and confidence levels. This study addresses the challenge of self-confidence estimation by comparing traditional machine-learning techniques with advanced deep-learning models. Our study involved a diverse group of thirteen participants (N=13), each of whom viewed and provided responses to seven distinct videos, generating eye-tracking data that was subsequently analyzed to gain insights into their visual attention and behavior. To assess the collected data, we compare three different algorithms: a Long Short-Term Memory (LSTM), a Support Vector Machine (SVM), and a Random Forest (RF), thereby providing a comprehensive evaluation of the data. The achieved outcomes demonstrated that the LSTM model outperformed conventional hand-crafted feature-based methods, achieving the highest accuracy of 76.9% with Leave-One-Category-Out Cross-Validation (LOCOCV) and 70.3% with Leave-One-Participant-Out Cross-Validation (LOPOCV). Our results underscore the superior performance of the deep-learning model in estimating self-confidence in video-based learning contexts compared to hand-crafted feature-based methods. The outcomes of this research pave the way for more personalized and effective educational interventions, ultimately contributing to improved learning experiences and outcomes.