Off-campus SJSU users: To download campus access theses, please use the following link to log into our proxy server with your SJSU library user name and PIN.

Publication Date

Spring 2026

Degree Type

Thesis - Campus Access Only

Degree Name

Master of Science (MS)

Department

Applied Data Science

Advisor

Guannan Liu; Mohammad Masum; Yue Luo

Abstract

Eye-tracking enables hands-free human–machine interaction, but gaze direction classification remains challenging due to noise and inter-user variability. This thesis presents a comparison of unimodal eye-tracking features using machine learning (ML) and deep learning (DL) models. Data were collected from five participants using a Pupil Core device across five gaze directions. Five features were extracted: temporal: raw gaze, fixation, and Ultimate Filtered Fixation (UFF); visual: heatmaps and eyeball images. Temporal features were evaluated using Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT), while visual features were evaluated using ResNet18 and MobileNetV2 under user-dependent and user-independent settings. In the user-dependent setting, fixation features achieved the highest ML performance (RF: 0.90, SVM: 0.88), followed by raw gaze feature and UFF. Heatmaps outperformed eyeball images in DL models (ResNet18: 0.74, MobileNetV2: 0.68 vs. 0.56 and 0.53). In the user-independent setting, fixation features remained dominant (SVM, GBDT: 0.94; RF: 0.93), while heatmaps showed strong generalization (MobileNetV2: 0.80, ResNet18: 0.79) and eyeball images remained lower (0.61, 0.55) than heatmap. These results establish that fixation features are most effective for ML, heatmaps for DL, and user-independent evaluation improves generalization across features and models.

Download

Available for download on Sunday, July 29, 2029

Share

COinS