Optimizing Temporal Convolutional Network for Eye Movement Classification Using Tree-Structured Parzen Estimator

Touchless technology has become increasingly vital for enhancing various aspects of human-computer interaction. One of these technologies is eye tracking. Eye tracking data processing comprises three phases: object selection, signal denoising, and event detection. In the event detection phase, eye movement is classified into three main classes: fixation, saccades, and smooth pursuit. Several studies have been conducted to improve the accuracy of eye movement classification. The latest research implemented the Temporal Convolutional Network that was optimized using Grid Search. However, the improvement in smooth pursuit accuracy is still lower than fixation and saccades. In this case, an accurate classification of smooth pursuit eye movement is vital to enhance spontaneous interaction in gaze-based touchless technology. To solve this research gap, we propose the Tree-structured Parzen Estimator (TPE) to optimize the Temporal Convolutional Network (TCN) for eye movement classification. Based on the experimental result, we achieved F1-scores of 94.50%, 89.83%, and 76.20% for fixation, saccades, and smooth pursuit eye movement, respectively. Comparative analysis with the state-of-the-art method shows that our method achieved higher improvement in smooth pursuit eye movement (0.40%) than fixation (0.11%) and saccades (0.22%). These results implies that the Tree-structured Parzen Estimator is promising to optimize time series deep learning for eye movement classification.