Figure 1.
Experiment procedure and stimuli.
(A) Flow chart of the experiment procedure. There were four consecutive stages in each trial: Mask (Gaussian-noise image), Stimuli (display of the target affective picture and monitoring of monocular eye position), Patch (patch image recognition task, to encourage free viewing) and Task (picture familiarity rating on a 7-point scale). (B) The scatterplot shows affective ratings of IAPS pictures in the valence block (VB, red triangles) and the arousal block (AB, blue squares) by females and males. Scores in the two dimensions were independent and decorrelated. (C) The bar graph shows valence ratings among three graded levels (HV, MV, LV) in VB (red bar), and arousal ratings among three graded levels (HA, MA, LA) in AB (blue bar). Three stars mark highly significant difference (p< 0.001, Kruskal-Wallis test).
Figure 2.
Path graph model of scan path.
(A and C) Schemas that illustrate two scan paths with loose and compact patterns. The shaded circles with numbers denote fixation sequence. Larger circles indicate longer fixation durations. The arrow line between two fixations indicates a saccade. (B and D) Weighted undirected path graph models corresponding to real eye movement scan paths in (A) and (C), respectively. The vertices represent gaze fixations. Each edge is weighted by the Mahalanobis distance between two adjacent vertices. (E) Spectral embedding in eigen-space. Two scan paths with different patterns were embedded into
eigenspace by the orthonormal basis
of its Laplacian matrix. (F) Spectral embedding in
eigenspace. Scan-paths were embedded into
space by the orthonormal basis
of its Laplacian.
Figure 3.
Shifts of pupillary reflex curves by valence and arousal.
The pupillary reflex curve is plotted by the pupil diameter of the first 15 consecutive fixations (duration ≈5 seconds) in the scan path and pooled over 22 subjects. (A) Shifts of the pupillary reflex curves by emotional arousal (p<0.01, repeated ANOVA). Pupil diameter while viewing low (LA) and high arousal pictures (HA) is larger than viewing medium arousal pictures (ML), indicating a non-linear effect of arousal (F(1,14) = 66.022, p<0.01). (B) Shifts of the pupillary reflex curves by hedonic valence (p<0.01, repeated ANOVA). Pupil size while viewing valenced pictures is linearly related to the valence level, suggesting a linear effect of valence (F(1,14) = 66.022, p<0.01).
Figure 4.
Topological metrics of scan paths.
(A) Fiedler constant for the three affective levels (H, M, L) of the valence and arousal blocks. The scan paths of arousal pictures have larger Fiedlers compared to scan paths of valenced pictures. Within the arousal block, the Fiedler is significantly different across three graded levels, indicating a non-linear effect of arousal. (B) Diameter of path graph. The average diameter of path graph for valenced pictures is larger than that for arousal pictures. Significant within-block differences could be observed in arousal block, indicating non-linear effect of arousal. (C and D) Energy metrics of and
embeding.
and
showed differences between the valence and arousal block. For arousal pictures, there is a significant within-block difference, suggesting a non-linear arousal effect. Star marks significant difference (one star: p<0.05; two stars: p<0.01 and three stars: p<0.001; by ANOVA test).
Figure 5.
Quadratic curves of spectral embedding.
The embedded path graph was fitted by the quadratic function in the vertex form of . (A) Coefficient a is positive, therefore all embedded quadratic curves opened upwards. Fitted quadratic curves in embedded path graph of valenced pictures have larger open width; the coefficient a is significantly different across three arousal levels, indicating a non-linear effect of arousal. (B) Coefficient b represents the symmetry axis of the fitted parabola. There was no significant difference between the valence and arousal blocks for coefficient b. (C) Coefficient c represents the height (vertical position) of the fitted quadratic curves. The fitted curves have larger coefficients c in valenced pictures and were significantly different within graded arousal levels, indicating a non-linear arousal effect. (D) Embedded quadratic curves in each affective level of two blocks. Stars mark significant difference (two stars: p<0.01 and three stars: p<0.001; by ANOVA test). Here
and
denotes the 2nd and 3rd eigenvector of the Laplacian matrix, respectively.
Figure 6.
Cubic curves of spectral embedding.
The embedded path graph was fitted by the cubic function in the form of
. (A - E) Coefficients
,
,
,
and
were compared between the two blocks and within each block. There are significant differences between valenced pictures and arousal pictures for the coefficients
,
,
and
, where coefficients
and
were different in graded arousal levels, indicating a non-linear effect of arousal. (F) The embedded curves in each affective level of valence and arousal. Star marks significant differences (one star: p<0.01; two stars: p<0.01 and three stars: p<0.001; by ANOVA test). Here
,
and
denote the 2nd, 3rd and 4th eigenvector, respectively.
Table 1.
Results of the multivariate linear model.
Figure 7.
Linear valence effect and non-linear arousal effect.
(A) Scan path found a significant linear trend across three graded valence levels (HV, MV and LV) in these metrics: edge weight (p = 0.02), peak velocity (PKV, p = 0.01), saccade angle (SAG, p<0.01) and pupil size (PUL, p<0.01), indicating a linear hedonic valence effect. (B) Scan path found a quadratic trend across three graded arousal levels (HA, MA and LA) in these metrics: fixation number (FNum, p<0.01), diameter of scan path (Diameter, p<0.01), coefficient a (p<0.01), coefficient c (p<0.01), PUL (p = 0.01), PKV(p<0.01), (p = 0.11),
(p = 0.13), weight(p = 0.07),
(p = 0.056),
(p<0.01), saccade duration (SDR, p<0.01) and fixation duration (DRA, p<0.01). By F-test.