|
Eye tracking technology is a critical component in human-computer interaction (HCI). For HCI systems that incorporate eye tracking, the precise classification of eye movement patterns is crucial. Yet, current deep learning methods face several difficulties such as data scarcity issue and class imbalance problem. To overcome these challenges, we introduce the ForestEye Classifier (FEC), which features a dual-component strategy: (1) Feature Extraction: The utilization of a multiscale time window approach enables the extraction of features spanning a variety of temporal scales. This method is crucial for the precise classification of different patterns, effectively addressing the challenge of data scarcity. Specifically, the classification of smooth pursuit relies on features from extended time scales. (2) Classification Model: FEC employs the deep forest ensemble learning method and utilizes a layered-ensemble architecture to integrate multiple forests, thereby overcoming the class imbalance problem and achieving improved classification performance. This marks a pioneering application of ensemble learning in the field of eye movement classification. In the end, the experiment indicates that, compared to the existing deep learning-based models, FEC demonstrates superior F1 scores (classification accuracy) across all eye movement patterns and within two distinct datasets. Moreover, FEC demonstrates consistent and robust performance across a range of participants.
|
