Eye Tracking in Recommender Systems

2025-12-18 · Source: Data Skeptic · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

Santiago de Leon, a researcher at the Kempelin Institute and Brno University, discusses the transformative potential of eye tracking in recommender systems. He explains how eye tracking captures gaze data, processing it into fixations and saccades to reveal user browsing patterns, which traditional click data often misses. De Leon introduces RecGaze, the first eye tracking dataset specifically for recommender systems research, enabling deeper understanding of user interaction with carousel interfaces like Netflix. His work, a collaboration between psychologists and AI researchers, uncovers insights such as positional bias and unexpected right-to-left browsing patterns after swiping. The discussion also addresses critical ethical considerations regarding pupil data and user privacy, emphasizing the importance of questioning assumptions in recommender systems and leveraging eye tracking to enhance user experience and de-bias algorithms.

Key takeaway

For AI Product Managers designing or optimizing recommender systems, integrating eye tracking insights can significantly improve user experience and recommendation accuracy. Your team should empirically validate assumptions about user browsing behavior, such as positional bias and swipe patterns, rather than relying solely on click data. This approach, potentially using datasets like RecGaze, allows for more informed interface design and algorithm adjustments, leading to more relevant and engaging recommendations, while carefully considering privacy implications.

Key insights

Eye tracking data offers richer insights into user behavior than clicks, enabling more effective recommender systems.

Principles

• Positional bias significantly influences user interaction.
• User browsing patterns can defy traditional assumptions.
• Eye tracking data is unique and highly personal.

Method

Eye tracking captures raw gaze data (XY position, timestamp), processed into fixations (>100ms) and saccades. This data is then aggregated and mapped to areas of interest (AOIs) to analyze browsing patterns and infer user intent.

In practice

• Consider flipping item ranking order after carousel swipes.
• Prioritize top two rows for critical content due to bias.
• Use eye tracking to build more realistic user simulators.

Topics

• Eye Tracking
• Recommender Systems
• RecGaze Dataset
• Positional Bias
• User Behavior Simulation

Best for: AI Scientist, Research Scientist, AI Product Manager, AI Researcher, Machine Learning Engineer, Data Scientist

Related on AIssential

• ST-DiffEye: Diffusion-based Continuous Gaze Generation via Joint Scanpath-Trajectory Modeling — ST-DiffEye jointly models gaze trajectories and scanpaths via diffusion to generate diverse, accurate human gaze patterns.
• How Data-Driven Grocery Recommendations Help Shoppers Eat Better With Less Effort — Data analytics and personalization are transforming grocery shopping into a data-driven decision-making process.
• Your Mouse and Eyes Secretly Leak Your Preference: LLM Alignment using Implicit Feedback from Users — Implicit user feedback, especially mouse movement, significantly enhances LLM alignment and preference prediction accuracy.
• Not All RecSys Problems Are Created Equal — RecSys complexity is dictated by catalog stability, signal strength, and preference subjectivity, not universal architectural mandates.
• Algorithms of Attraction: The Digital Cupid Within Modern Dating Apps — Modern dating apps increasingly leverage sophisticated AI algorithms to move beyond static filters, dynamically predicting compatibility and optimizing the matching experience.
• Stop tracking hypothetical prompts: how to build a prompt set from real buyer queries — Accurate AI visibility tracking requires prompt sets built from real buyer queries, not speculation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Data Skeptic.