Fig 1.
Experimental Setup and Seahaven.
(a) The participant sits on a fully rotatable swivel chair wearing headset and controller. The experimenter’s screens in the background allow monitoring the participant’s visual field of Seahaven (left) and the pupil labs camera images (right). (b) The island of Seahaven in aerial view.
Fig 2.
(a) We interpolate missing data only if less than eight samples are missing consecutively (pupil detection with less than 50% probability) and if these samples occur between two clusters on the same collider. During the interpolation process, these samples are then unified with the two clusters to form a new cluster on the same collider. The first row shows three clusters of missing samples (marked noData), while the second row represents the result of the algorithm. In the first cluster (green box), both interpolation conditions apply: there are fewer than eight consecutive missing data samples (#4) and they are surrounded by two clusters on the same collider. Consequently, these missing samples are interpolated and combined to a new cluster. In the second cluster (1st orange box) the first interpolation condition applies (#noData samples < 8) but the cluster occurs between clusters on two different colliders (H103 and H54). Therefore, no interpolation is performed. In the third cluster (2nd orange box) only the second interpolation condition applies. Even though the missing data samples occur between two clusters on the same collider (H55), the first interpolation condition is violated (#noData samples ≥ 8). Consequently, no interpolation is performed. (b) Histogram of hit point cluster length distribution after interpolation. The distribution is visualized between 0 and 1033 ms. The longest hit point cluster on a house has a duration of 18.9 seconds. The ordinate corresponds to the probability of each duration. Since previous work used gamma distributions to model the distribution of fixation durations or response latencies [39,40], we model the two partly overlapping gamma distributions for visualization only, fitting the distributions of the duration of fixations (green) and non-fixation events (grey). The dashed red line marks the separation threshold for gazes. (c) The pie chart shows the result of the gaze classification across all participants.
Fig 3.
(a) Timeline of gaze events by a participant. The abscissa represents the first 30 seconds (900 hit points) of the recordings. The ordinate contains all viewed houses viewed during that time line. We number houses and name them accordingly, e.g., H148 for house number 148. In this panel each house has a distinct color for visualization only. The grey bars represent clusters of the NH category, which are not considered during graph creation. The black bar identifies a remaining cluster of missing data samples. Therefore, no edge will be created at this moment in the timeline. (b) The graph corresponding to the timeline of panel A is visualized on top of the map of Seahaven. The colors of the nodes match the colors of the boxes in panel A. Edges are labelled according to the order they were created. (c) The complete graph of a single participant based on all gaze events during 90 min of exploration visualized on top of the 2D map of Seahaven. Note that in this visualization the locations of the nodes correspond to the locations of the respective houses they represent in Seahaven, however, this locational information is not contained in the graph itself.
Fig 4.
(a) The sparsity pattern of the graph’s adjacency matrix sorted by the entries in second smallest eigenvector. Color coded into edges between nodes of one cluster (green), edges between nodes of the other cluster (blue), edges between nodes of the two clusters (black) and a distinction between the clusters (yellow). (b) The two clusters are displayed onto the map of participant 35. (c) The second smallest eigenvector of the Laplacian matrix is sorted ascendingly and color coded into two clusters.
Fig 5.
(a) The graph of one participant is visualized on top of the map of Seahaven. The nodes were colored according to their respective node degree centrality. (b) The individual mean node degrees of all participants (blue line) and their respective standard deviation (grey lines) are depicted sorted by increasing values. (c) A pseudo 3D plot color coding the node degree of every house (abscissa) for every participant (ordinate). The houses are sorted so that the average node degree value increases along the abscissa. Similarly, the participants are sorted, so that the average node degree increases along the ordinate. The marginals of this plot result in the panels b and e. (d) The distribution of the pairwise inter-participant correlation coefficients of the node degree values of all houses. (e) The mean node degree of each house sorted according to the mean node degree along the abscissa (blue line) and their respective standard deviations (grey lines).
Fig 6.
(a) The frequency of occurrence of the node degree for a single participant. The green line indicates the linear regression starting at the median of the distribution. (b) The distribution of the hierarchy index across all participants.
Fig 7.
The high node degree centrality houses.
(a) The mean node degree distribution across all participants with mean, 1σ- and 2σ-thresholds. (b) A box plot of the 10 houses with at least 2σ-distance to the mean node degree. (c) A map plot with all nodes color coded with their respective average node degree across all participants. (d) The 10 houses, which exceeded the 2σ-distance to the mean displayed on the map for our example participant with all their connections and color coded with their respective node degree.
Fig 8.
(a) The development of the rich club coefficient with increasing node degree. The dot-lines are the rich club coefficients of individual participants, while the green line is the mean across all participants. (b) All houses displayed on the map both color coded and size coded according to their frequency of being part of the rich club across participants.
Fig 9.
Location data of all participants plotted on the map of Seahaven. The color code indicates how many of the gaze-graph-defined landmark houses were viewed by participants at each location.
Fig 10.
Example buildings in Seahaven.
a) The upper two rows show the ten houses being labeled as gaze-graph defined landmarks. House 1 being the house with the highest and house 2 with the second highest node degree. b) The lower two rows show ten other randomly selected houses out of the remaining 203 houses that are located in Seahaven.
Fig 11.
(a) The node degree distribution of an example graph with similar distribution to the graphs used in the publication. (b) The example graph with color-coded nodes according to their individual node degree defined by the number of edges that connects a node to other nodes.
Fig 12.
Graph partitioning–Toy example.
(a) The sparsity pattern of the graph’s adjacency matrix sorted by its second smallest eigenvector. Color coded into edges between nodes of one cluster (blue), edges between nodes of the other cluster (green), edges between nodes of the two clusters (black) and a distinction between the clusters (yellow). (b) The toy graph with two color coded clusters. (c) The second smallest eigenvector of the Laplacian matrix is sorted ascendingly and color coded into two clusters by division into positive and negative components.
Fig 13.
(a) The node degree distribution of the example graph. (b) The example graph (c) A scatter plot of the logarithm of degree against the logarithm of the frequency of occurrence. The green line shows the linear fit to the part of the data above the median degree.
Fig 14.
Rich club coefficient–Toy example.
(a) The node degree distribution of the example graph. (b) The example graph with color-coded nodes according to their node degree. (c) The adjusted rich-club coefficient, i.e., normalized using a random graph with similar node degree centrality statistics.