Continuous ratings of movie watching reveal idiosyncratic dynamics of aesthetic enjoyment

Abstract

Visual aesthetic experiences unfold over time, yet most of our understanding of such experiences comes from experiments using static visual stimuli and measuring static responses. Here, we investigated the temporal dynamics of subjective aesthetic experience using temporally extended stimuli (movie clips) in combination with continuous behavioral ratings. Two groups of participants, a rate group (n = 25) and a view group (n = 25), watched 30-second video clips of landscapes and dance performances in test and retest blocks. The rate group reported continuous ratings while watching the videos, with an overall aesthetic judgment at the end of each video, in both test and retest blocks. The view group, however, passively watched the videos in the test block, reporting only an overall aesthetic judgment at the end of each clip. In the retest block, the view group reported both continuous and overall judgments. When comparing the two groups, we found that the task of making continuous ratings did not influence overall ratings or agreement across participants. In addition, the degree of temporal variation in continuous ratings over time differed substantially by observer (from slower "integrators" to "fast responders"), but less so by video. Reliability of continuous ratings across repeated exposures was in general high, but also showed notable variance across participants. Together, these results show that temporally extended stimuli produce aesthetic experiences that are not the same from person to person, and that continuous behavioral ratings provide a reliable window into the temporal dynamics of such aesthetic experiences while not materially altering the experiences themselves.

Fig 1. Schematic description of measuring continuous aesthetic responses to dynamically changing visual experiences.

(A) Participants viewed videos of landscapes or dance performances for 30 seconds while making continuous ratings of the moment-to-moment enjoyment they were having. This was immediately followed by an overall rating indicating the intensity of their aesthetic experience of the clip. Both types of responses were made using a dial that controlled the slider display on the screen. (B) The rate group completed continuous and overall ratings in both test and retest sessions. (C) The view group gave overall ratings but not continuous ratings in the test session. In the retest session, participants in this group performed both types of ratings.

Fig 2. Distributions of mean overall ratings across groups (rate vs. view), sessions (test-retest) and categories (dance vs. landscape).

On average, landscape videos were rated higher than dance videos, but there was no main effect of session or group (The diamond symbols show the means and black vertical lines indicate 95% confidence intervals of the means).

Fig 3. Representative continuous rating traces.

(A) Continuous rating traces from one participant for each movie clip. In addition to differences in overall mean liking, some movies elicited more temporal dynamics than others. Rows show different categories (Dance, Landscape) and columns show ratings for different sessions. (B) Continuous rating traces given by each participant for one dance and one landscape clip. These movie clips were rated remarkably differently by different participants, both in terms of its overall mean liking, but also in terms of variability over time. Each row shows a sample movie from one genre (Dance, Landscape) and columns show ratings for different sessions.

Fig 4. Test-retest reliability of continuous ratings.

The distributions of the reliability values both with (A) Pearson correlations and (B) L_2-norm values indicate that most participants showed good test-retest reliability with no difference across categories. (C) There was positive correlation between L_2-norm values for overall ratings and median L_2-norm values for continuous ratings per participants.

Fig 5. Temporal variability in the continuous ratings.

(A) Distributions of mean rmsd values for participants are quite similar across sessions and categories and generally higher than the rmsd value obtained from monotonic increase (inline panel) over the entire clip (red dashed lines). (B) 95% prediction intervals for the random effect estimates of 30 movie clips and 50 participants show that the amount of temporal variability differed more across participants than movies. (Note different x-axes scales for panels.) (C) K-means clustering analysis with median rmsd values resulted in 2 different clusters of participants (shown with green and black). Note that the participants in the first cluster (green-fast responders) have the highest random effects conditional modes.

Fig 6. Agreement across participants with “mean-minus-one” (MM1) and variance decomposition.

(A) Overall rating agreement with MM1: Rate group’s agreement decreased in the retest session whereas view group did not show this session effect. (B) Continuous rating agreement with MM1: Agreement is higher for landscape videos for both groups and retest session agreement is lower for the rate group. (Error bars show 95% confidence intervals calculated with the z-transformed r values, horizontal dashed line indicated zero) (C) The proportion of repeatable variance that is attributable to individual taste is higher than the proportion of variance that is attributable to shared taste for both categories and groups.