A gender-aware saliency prediction system for web interfaces using deep learning and eye-tracking data.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Cognitive Science

Background:

• Understanding demographic influences on visual attention is vital for user-centered web design.
• Previous research has not fully explored gender-specific gaze behaviors in web browsing.

Purpose of the Study:

• To develop a gender-aware saliency prediction system using deep learning.
• To investigate gender-related differences in visual attention patterns on web pages.
• To introduce and utilize the WIC640 dataset for demographic-aware AI research.

Main Methods:

• Fine-tuning TranSalNet, a Transformer-based saliency prediction model.
• Utilizing the WIC640 dataset with 640 web page screenshots and eye-tracking data from 85 participants.
• Analyzing gaze behavior patterns and comparing gender-specific models with a general model.

Main Results:

• Distinct visual saliency patterns were observed between male and female users.
• Gender-specific models showed performance variations (e.g., female-trained: CC=0.7786, NSS=2.4224; male-trained: CC=0.7582, NSS=2.3508).
• A general model trained on inclusive data outperformed gender-specific models, indicating the importance of diverse training datasets. Significant gender differences were found in 9/12 saliency features, and age correlated with reduced fixation dispersion.

Conclusions:

• While gender influences visual attention, inclusive AI training data is crucial for optimal performance in web interface design.
• The findings suggest potential for personalized user experiences through demographic-aware AI systems.
• The WIC640 dataset provides a valuable resource for future research in adaptive AI, visual attention, and interface design.