Krzysztof Krejtz, Tomasz Szmidt, Andrew Duchowski, Izabela Krejtz
10 min
Quantifying and comparing eye movement transitions between Areas of Interest (AOIs) remains a challenge in gaze analysis, as existing metrics like scanpath edit distances and heatmap overlaps either ignore sequential order or lack robust statistical significance testing. This paper introduces a two-step method to quantify individual gaze switching patterns. First, an individual's sequence of fixated AOIs is modeled as a first-order Markov chain, validated using goodness-of-fit tests. Second, Shannon's entropy coefficient is calculated from the individual's transition matrix to measure the complexity and predictability of gaze switching (transition entropy), while the entropy of the stationary distribution measures how evenly visual attention is spread across different AOIs.
The authors demonstrated this method using eye-tracking data from 43 participants who freely viewed classical art paintings from Impressionist, Renaissance, and Bauhaus periods. Before viewing the artwork, participants completed the Curiosity and Exploration Inventory and underwent a mood induction procedure. Afterward, they rated the attractiveness of each painting. Goodness-of-fit tests confirmed that the first-order Markov chain assumption adequately modeled individual AOI transition sequences for the vast majority of participants across different numbers of defined AOIs.
The entropy metrics successfully captured meaningful individual differences in visual exploration. Transition entropy followed a Gaussian distribution across subjects, whereas stationary entropy distributions deviated from normality, indicating that nonparametric tests should be used for stationary entropy analyses. Qualitative comparisons of scanpaths and transition matrices showed that individuals with high transition entropy frequently switched between AOIs, whereas those with low transition entropy engaged in more sustained, localized viewing within the same region. Similarly, varying stationary entropy scores reflected whether a viewer concentrated attention on specific objects or distributed it evenly across the entire scene.
Sam: The paper addresses this directly. They ran statistical tests comparing the real gaze data against what the Markov model would predict. For the vast majority of participants and viewing sequences, the fit was good — meaning the assumption that your next look depends mainly on your current location, not your whole history, is a reasonable description of how human eyes actually move.
Alex: That's a useful sanity check. So once the model is validated, what do the entropy numbers actually reveal?
Sam: A few things. When researchers increased the number of distinct zones on the screen — say, from two regions to three — the average entropy went up. That makes intuitive sense: more zones means more possible paths between them, so the system becomes harder to predict by default. They also found differences between artworks. A Renaissance painting produced higher transition entropy than Impressionist or Bauhaus works, which the authors suggest reflects the more complex narrative composition of that particular piece — more distinct focal points pulling the eye in different directions.
Alex: And what about the connection to how much someone actually liked what they were looking at?
Sam: This is one of the more interesting findings. Participants who gave an artwork higher ratings tended to show lower transition entropy — meaning their eye movements were more structured and deliberate. The authors suggest that when people genuinely appreciate something, they explore it in a more focused, orderly way, rather than scanning restlessly.
Alex: So a more chaotic scanpath might signal disengagement, while a more structured one suggests genuine interest.
Sam: That's the interpretation the paper offers, with appropriate caution — it's a correlation observed in this study, not a universal rule. And it's worth being clear about the limitations here. The whole method depends on how researchers define the zones on the screen in the first place. Draw too many small zones, and you end up with a data matrix full of empty cells — the eye simply never jumps between most pairs of zones — which makes the entropy estimates unreliable. The authors recommend keeping the number of zones modest and ensuring participants view the image long enough to generate sufficient data.
Alex: So the quality of the output is only as good as the care taken in setting up the zones at the start.
Sam: Exactly. It's a meaningful methodological step — turning something as subjective as visual wandering into a rigorous, comparable number — but it requires careful design choices to work well. The hope is that future work will develop clearer conventions for how to define those zones consistently across studies.
Alex: That's a satisfying arc — from the problem of measuring something as personal as where someone looks, to a principled statistical framework for doing exactly that. Thanks for walking through it, Sam.
Sam: It's a useful contribution to how we connect eye-tracking data with cognitive and aesthetic experience. Thanks for listening to ResearchPod.