Visualization Lenses for Time-Binned AOI Occupancy in Eye-Tracking Analysis
Contribution
Frames aggregate AOI analysis as a small design space: the questions asked of cohort gaze data reduce to three task families—reading magnitude, judging which AOI leads in each bin, and spotting matrix-level structure such as sparsity or synchrony—which, by graphical-perception theory, no single chart serves well.
Characterizes four established chart types (Stream, Ridgeline, Distribution, Heatmap) by their trade-offs for AOI data rather than as novelties, naming where each misleads: the streamgraph’s moving baseline, the lost part-to-whole sense of separated ridgelines, the broken object permanence of per-bin reordering, and color’s weak channel for magnitude.
Keeps one deliberately simple quantity behind all four views—the share of the cohort fixating each AOI per time bin—so changing the chart never changes the numbers. The contribution is the discipline of one substrate, four projections; not the (elementary) quantity itself.
Couples the aggregate views with participant-level scarf plots in one workspace, so an analyst moves from a cohort pattern straight to the individuals behind it—overview to detail—without re-exporting data or switching tools.
Supports reproducible exploration by exporting the full analytic state—data, layout, filters, parameters, tool version—so a configuration can be reopened and shared, closing the documentation gap that makes no-code tools hard to reproduce.
Treats honest scoping as part of the contribution: it states what aggregation discards (scanpath order, transitions, individual differences—so a split cohort can look uniform) and what the views are sensitive to (bin width, AOI definition), positioning them as hypothesis-generating rather than confirmatory.
Publication properties
Citation
Vojtechovska, M., Beitlova, M., & Popelka, S. (2026). Visualization Lenses for Time-Binned AOI Occupancy in Eye-Tracking Analysis. In Proceedings of the 2026 Symposium on Eye Tracking Research and Applications (ETRA 2026). ACM. https://doi.org/10.1145/3797246.3805723
Authors
Year
2026
Journal
Proceedings of the 2026 Symposium on Eye Tracking Research and Applications
Original language
EN
Abstract
Questions addressed
Q: How do you summarize where a whole group looked over time, when a scarf plot only shows one person?
A: You move from individual traces to a cohort summary: divide the recording into equal time bins and record what share of the group fixated each area of interest in each bin. Plotting that share over time shows how collective attention shifts—shared phases, hand-offs between elements. The price is the loss of each person’s exact scanpath; order and transitions stay in the individual scarf plots, so the two are meant to be read together.
Q: Why use several different charts for the same aggregated eye-tracking data instead of one?
A: Because the questions differ and, by graphical-perception research, no single encoding answers them all well. A stacked silhouette captures overall engagement but reads one AOI’s values poorly; separated curves read precisely but hide the part-to-whole picture; rank, dominance, and sparsity need yet other layouts. Several views of the same numbers let the analyst match the encoding to the question instead of forcing one chart to do everything.
Q: Stream, Ridgeline, Distribution, Heatmap—which question does each one answer best?
A: Stream (a centered streamgraph) is the overview: overall engagement and big shifts in how attention is split. Ridgeline (separate curves on a shared scale) follows one AOI and shows when it peaks. Distribution (bars re-sorted by size each bin) shows which AOI leads and when the lead changes. Heatmap (a color grid of AOI × time) handles many AOIs at once—empty stretches, moments when several light up together—where height- or area-based charts get too crowded.
Q: What does aggregate AOI visualization throw away, and when does that matter?
A: Summing across people discards scanpath order and transitions, and—most importantly—averages away individual differences: a cohort split into two strategies (half fixate an element intensely, half ignore it) can look like moderate, uniform interest. Results also depend on choices made beforehand—bin width and how the AOIs were drawn. So treat these views as a way to generate hypotheses, then check them against individual data and alternative settings.
Q: Why can a cohort’s attention add up to less than 100% at a given moment?
A: Only fixations inside a defined AOI count toward the total. Saccades, blinks, lost tracking, looks at unmapped regions, and participants who already finished contribute nothing, so a bin can sum to well under 100%. The views keep this visible rather than hiding it—in the Stream the silhouette thins and tapers toward the end—because it carries real information about saccadic phases, data quality, and who is still on task.
Q: If streamgraphs and heatmaps already exist, what is actually new here?
A: Not the charts, and not the underlying number—the share of a group fixating each AOI per bin is deliberately simple. What the paper adds is the reasoning around them: a clear map from question to encoding, an honest account of where each encoding misleads, and the integration of all four with individual scarf plots and exportable, reproducible state in one open-source tool. The value is the design knowledge and the workflow, not the novelty of the visual forms.