Eye Tracking in Recommender Systems

Eye‑tracking technology captures a user’s gaze as an XY coordinate stream, typically sampled at 90 Hz or higher. Researchers convert raw gaze points into fixations—moments longer than 100 ms that indicate cognitive processing—and saccades, the rapid jumps between fixations. By overlaying these fixations on interface elements such as movie posters, analysts can quantify exactly which items attract visual attention and for how long. In recommender‑system research this richer signal goes beyond traditional click‑through data, offering a near‑real‑time window into subconscious preferences. The method, however, demands careful mapping of screen coordinates to defined Areas of Interest (AOIs) and substantial preprocessing to become human‑readable.

The first large‑scale eye‑tracking study on Netflix‑style carousels revealed several surprising browsing patterns. Participants consistently began scanning from the far right of a row after swiping, then moved leftward, contradicting the classic left‑to‑right reading assumption. The top two rows attracted a pronounced fixation bias, while complete row skipping was rare, indicating users explore most presented categories. Interestingly, the second‑to‑last item on a row received slightly longer dwell time than the final slot, suggesting subtle positional effects within a single carousel. These insights imply that ranking algorithms should consider reversing or rotating item order on subsequent swipes to align with natural eye movement.

Integrating fixation data into recommender pipelines can dramatically refine positional bias models that currently rely on clicks or impression logs. Eye‑tracking distinguishes between items merely displayed and those actually viewed, enabling more accurate estimation of true user interest and reducing the noise inherent in click‑only feedback. For businesses, this translates into higher conversion rates, better ad placement, and more personalized content curation. Future work may compare visual versus positional bias across cultures, especially right‑to‑left reading markets, and combine gaze data with saliency‑prediction networks to automate AOI detection. As mobile devices embed eye‑tracking sensors, the opportunity for real‑time, privacy‑preserving recommendation enhancements grows rapidly.