Network modeling of aesthetic effect for Chinese Yue Opera costume simulation images

3.1 Aesthetic cognitive and its empirical study

The field of aesthetic cognitive psychology has experienced significant advancements. Leder et al. [6] proposed the renowned Model of Aesthetic Appreciation and Aesthetic Judgment. This model aims to elucidate why individuals are attracted to artworks and explore how aesthetic cognitive processing elicits aesthetic emotions and overall positive and self-rewarding aesthetic experiences. According to the model, aesthetic appreciation and judgment encompass five distinct stages: perceptual analysis, implicit memory integration, explicit memory integration, cognitive mastery, and evaluation. The outcome of aesthetic processing includes aesthetic emotions and aesthetic judgments. Thus, the cognitive mastery and evaluation stages of aesthetic appreciation and judgment establish a feedback loop [31]. For the visual art, Chatterjee proposed an overall framework of the neural basis of visual aesthetics [32]. This framework emphasizes the transition from spontaneous to self-aware evaluation. In 2014, Chatterjee also proposed the aesthetic triad structure. He believed that aesthetic experience includes the emotions, evaluations, and behaviors caused by these things, as well as the processes that constitute their interpretation and products [33].

Empirical research on aesthetic cognitive experience primarily focuses on the public’s aesthetic evaluation of specific artworks [34]. In the quantitative research on aesthetic cognitive experience, except for the eye-tracking experiments [10,34], domain-specific aesthetic scales have been created. For instance, some studies have utilized semantic bipolar scales based on Osgood et al.’s EPA model (EPA represents three dimensions of semantic evaluation: Evaluation, Potency, and Activity) [35]. The Survey for the Assessment of Aesthetic Perception (SAAP), proposed by Rowold in 2008, includes 16 factors encompassing emotion, cognition, and self-consistency. However, the SAAP does not fully cover the affective and cognitive processes related to beauty [36]. Hager et al. [37] introduced the Art Reception Survey Scale (ARS) in 2012, which utilizes a five-level Likert scale and consists of six factors: cognitive stimulation, negative emotion, professional knowledge, self-reference, artistic quality, and positive attractiveness, comprising a total of 30 items. This questionnaire has been tested and applied extensively. There are also the Aesthetic Emotion Scale developed by Schindler et al. [38] and the Aesthetic Experience Questionnaire (AEQ) developed by Wanzer et al. [39]. These studies provide references for scale development in this study.

3.2 Network analysis and its applications in psychology

Network analysis originated in the 1930s and began to be applied mainly in sociology [40]. With the increasing emphasis on the influence of social situations in psychological research [41,42], network analysis methods have been applied to the psychological research. The most widely used fields of this method are clinical psychology, psychometrics, and personality research. For example, psychiatric research attempted to use network models to understand and predict the dynamics of psychopathology [43,44,45]. Epskamp et al. [46] introduced the network model as a formal psychometric model, conceptualizing the covariance between psychometric indicators as resulting from pairwise interactions between observable variables in a network structure. Costantini et al. [47] analyzed individual differences (e.g., in personality and psychopathology) and their patterns of stability and variability via network analysis.

Recently, Specker et al. [19] introduced network models to explore the aesthetic domain, offering new perspectives on conceptualizing aesthetic experience. They proposed an AEN, which highlights the direct relationships and mutual influence between aesthetic effects. Drawing upon network communication theory [48,49], Specker et al. constructed a statistical network to model associative processes between aesthetic effects. In this network, nodes represent experiences of aesthetic perception, and edges represent connections between these nodes. Empirical experiments conducted with abstract artworks confirmed their hypothesis and validated the AEN model to a certain extent. Importantly, these findings did not invalidate the results of the EPA model but rather served as a complementary addition to the previous research. The psychological network changes from representing observations to quantitative nodes (e.g., mood states, symptoms, or attitudes) and connections representing statistical relationships, and can be passed to calculate the centrality index of the node and the weight of the link to evaluate each variable interrelationship [50,51].

In a psychological network, partial correlation coefficients are commonly used to assess the connections between nodes. The Least Absolute Shrinkage and Selection Operator regularization algorithm [52] was often chosen to exclude sampling errors in the network [53]. During network visualization, the width of the line and the saturation of the color directly reflect the magnitude of the bias correlation coefficient: the wider the line and the more saturated the color, the larger the bias correlation coefficient [54]. The centrality of a node reflects the position of the node in the network, and it is common to use node strength, closeness, betweenness, and expected influence as four indicators to reflect the size of the centrality of a node. Some scholars are convinced that for psychological networks, betweenness and closeness centrality seem to be particularly unsuitable as measures of node importance [55]. As an alternative, the expected influence is the sum of all edge values connected to a particular node, indicating that the point is more important in the network [56].

In this study, the network analysis was exploratively applied to investigate the aesthetic experiences associated with Yue Opera costumes. Network research helps to reflect patterns of coexistence between variables of interest to the researchers. In addition, finding influential nodes and manipulating them help to practically improve the nodes associated with them throughout the network.

4.1 Stimuli

To investigate audience’s aesthetic preferences and perceptions of Yue Opera costume virtual simulation images, five images representing common traditional Yue Opera costumes were created using clo3D 7.0. These images were based on the virtual simulation methods for Yue Opera costumes as described by Liu et al. [1] and Han [22]. We utilized 2D virtual images instead of direct 3D models. 2D virtual simulation images are the representations of a specific viewpoint of the 3D virtual models. We did not let the subjects directly interact with the models because their attention might be drawn to various interactive forms, which could potentially impact their aesthetic appreciation of the costumes themselves. In addition, 2D virtual simulation images have advantages over photos of actual costumes in terms of highlighting detailed features, offering control over lighting and background settings, and ensure consistency in experimental conditions. The costumes depicted five different roles: Huadan, Shusheng, Laosheng, Zhengdan, and Wusheng. Each material had a resolution of 1,752 × 1,023 pixels, presenting the costumes from the front and back perspectives (as illustrated in Figure 1).

Figure 1

Virtual simulation images of Yue opera costumes for different roles. (a) Huadan (young females), (b) Shusheng (young males), (c) Zhengdan (primarily middle-aged females), (d) Wusheng (young and middle-aged male martial artists), and (e) Laosheng (senior males).

4.2 Aesthetic effect

To assess the aesthetic effects of Yue Opera costumes, a set of 16 items was included in the measurement. We expected to see the influence of objective elements of Yue Opera costumes on the whole aesthetic experience through the network modeling. Therefore, unlike the work of AEN, our measurement table contained a wider range of objects. From the order of abstraction from low to high, these items consist of three items that measured the audience’s perception of the low-level design elements (L1–L3), five items of semantic difference (S1–S5), and eight items of abstract aesthetic experience (A1–A8) (Table 1). Specifically, the liking for underlying design elements measures the evaluations of the color, pattern, and shape of the garment. The semantic differential measurement included five groups of semantic scales, with each group consisting of semantically opposite word pairs, i.e., sober–showy, elaborate–simple, dim–conspicuous, low level–high level, light–heavy, as referenced from Han’s research findings [57]. Here, we aimed to incorporate semantic differential measurements to gauge the audience’s perceptual imagery based on the basic design elements. In addition, we contextualized the measures based on scales such as the ARS [37], the Aesthetic Emotion Scale [38], and the AEQ [39], which were originally developed to assess the aesthetic experience of different artworks. Compared with L1–L3 and S1–S5, A1–A8 described more abstract aesthetic sensations, such as emotions, interests, understanding of costumed characters, and so on. All measurement scales were presented in the form of a seven-point scale, allowing participants to freely rate their evaluations of Yue Opera simulation images. The inclusion of these scales and experimental stimuli formed a 16-item experimental design, enabling a relatively comprehensive examination of the audience’s aesthetic responses to Yue Opera costumes.

4.3 Participants

Our experimental subjects were mainly focused on the youth group. The primary reason for choosing this group is the aging audience of Yue Opera in real life, with most being over 50 years old, leading to a potential generational gap in its inheritance. To investigate the visual aesthetic preferences of younger audiences toward Yue Opera, our study specifically targeted this group. The questionnaire was administered on the NAODAO online research platform [57] (www.naodao.com), and the participants consisted of registered members of the platform who were mainly youth in China. The experiment followed the ethical norms of the platform. A total of 257 participants were recruited, including 80 males, 170 females, and 7 who were not willing to disclose their gender. The mean age of the participants was 26.06 years with a standard deviation of 7.33 years. Among the participants, 159 individuals were aged 18–25 years, comprising 61.87% of the sample; 34 were aged 25–30 years, accounting for 13.23%; and 64 were older than 30 years, making up 24.90%. They were given informed consent before the participation.

4.4 Procedure

The experiment was developed using PsychoPy [58] version 2022.2.4 and conducted on the NAODAO online research platform. The detailed procedure is outlined in Figure 2.

Figure 2

Experimental procedure.

Before the start of the experiment, participants were provided with an instruction guide explaining the procedure. This ensured that participants had a clear understanding of the evaluation criteria, which means score should be based on current viewing experience, and there was not a right or wrong answer. The experiment consisted of 80 questions (5 pictures depicting different characters in Yue Opera costumes × 16 items). Participants used their mouse to rate each question on a seven-point scale, ranging from 1 to 7 including decimals. There was a 500 ms interval before each image appeared, and participants were not restricted by time limits when answering the questions. To maintain validity, data from subjects whose total participation time were less than 3 min were excluded. After verifying the collected data, participants were compensated with a subject fee as a token of appreciation.

4.5 Network estimation

To estimate the network of aesthetic effects, we referred to the analysis source code of Specker et al. [19], which followed the approach of Rhemtulla et al. [49] and Fried et al. [50]. All analyses were conducted in R using the “qgraph” package [59] within a Gaussian graphical model (GGM) framework. The GGM represents a network in which the edges connecting aesthetic effects estimate partial correlations, incorporating graphical lasso regularization. In the GGM, the edges can be interpreted as conditional dependencies between aesthetic effects. If two aesthetic effects are connected in the outcome graph, it indicates a dependency between them after controlling for all other variables. Conversely, if no edges appear, the aesthetic effects are conditionally independent. Regularization involves estimating a statistical model with an additional penalty for model complexity [53]. The global network is an average over the estimates of the individual networks for each costume. The gamma was set as 0.5.

To test the stability, the CS coefficients of the expected influence for five individual networks were calculated and all above 0.5, suggesting a relative estimated robustness. The stability of central indices, bootstrapped confidence intervals of estimated edge-weights, and the figures showing if the edges significantly differ from each other (nBoots = 2000) are shown in Figure A1.

5.1 Descriptive statistics

Figure 3 presents the mean scores for each scale. A two-way analysis of variance (ANOVA) indicated that the subjects’ scores differed significantly on costume (F = 38.700, p = 0.000, η² = 0.131), item (F = 278.118, p = 0.000, η² = 0.521), and their interaction (F = 57.632, p = 0.000, η² = 0.184). Table 2 shows the detailed effects of the dress type on subjects’ scores on each item. It can be seen that negative emotion (A2) has the lowest score among the five networks. L1–L3 (liking of pattern, color, and shape) and S1–S5 (showy, simple, conspicuous, high level, heavy) had relatively large differences in the scoring results. In comparison, the scores of A1–A8 (high-level aesthetic scales) had almost similar trends.

Figure 3

Average score on the aesthetic effect scale.

In addition, we categorized our samples into male and female groups to analyze the impact of gender differences on aesthetic experience. Most aesthetic preferences showed no significant distinctions between genders. However, for four specific questions – namely, “low level-high level,” “I want to know the cultural knowledge about this costume,” “I know what the creators are trying to convey with this costume,” and “This costume reminds me of certain life experiences from the past” – male responses scored slightly higher than female on average (t = −2.028, p = 0.044; t = −2.351, p = 0.020; t = −2.712, p = 0.007; t = −4.691, p = 0.000). This subtly suggests that females may have higher standards for the quality of Yue Opera costumes and are more cautious in their understanding of the cultural knowledge and creative intentions behind these costumes.

5.2 Global network

We first explain the overall network of Yue Opera costume aesthetics (averaged across all five individual networks), as shown in Figure 4, which represents a preliminary explanation of the relationship between aesthetic structures. The blue line connection represents the positive correlation, and the red line connection represents the negative correlation. The wider the line, the larger the partial correlation coefficient. According to the prompt of “qgraph”, care needs to be taken in interpreting very small relationships since a dense regularized network was selected, and we set edges with weights below 0.05, which is not shown in the figure. In this network, there was a strong positive correlation between A1: archetypal aesthetics and cloth L1: pattern (0.226), L2: color (0.209), L3: shape (0.116), A3: positive emotion (0.140), S4: high-level (0.137), and A5: interesting (0.122). Overall, based on the relative distance of the nodes in the network, they can be roughly categorized into three groups: the first group (L1–L3, A1–A5, and S4) where archetypal aesthetic was the center, the second group (A6–A8) representing self-reference and professional knowledge, and the last group (S1–S3 and S5) representing semantic evaluations expect for the sense of high-level. Among the groups, conspicuous (S3), creative intent (A7), knowledge (A4), positive emotion (A3), and shape (L3) could be seen “bridges” that triggered the connection of the three groups to a certain extent.

Figure 4

Global network of aesthetic effects across all Yue Opera costumes.

Considering that the grouping was based on networks without small edges, we further utilized unregularized GGM modeling (ggmModSelect) to calculate and generate the global network again. According to Figure A2, a similar layout can still be derived.

Like the analysis of AEN, we calculated the expected influence of each node (Figure 5a). Averagely, A1: archetypal aesthetics had the highest expected influence in the network (1.046), followed by the liking of L1: pattern (0.667) and L2: color (0.827). The liking of L3: shape (0.453), S4: high-level (0.406), A3: positive emotion (0.439), A4: knowledge (0.631), A5: interesting (0.496), A6: character (0.461), and A7: creative intent (0.481) had similar expected influence. A2: negative emotion (−2.208) had the lowest expected impact, followed by S2: simple (−1.776) and S5: heavy (−1.632).

Figure 5

Expected influence of nodes of the global network (a) and individual networks (b).

5.3 Network for different characters

Since the network structure may be different due to character creations, we illustrated the Yue Opera costume aesthetic network for each different character in Figure 6. Edges with weights below 0.1 were not shown. It can be seen that the basic network structures showed a strong similarity despite the larger differences in the ratings caused by the costumes. For example, there was a positive correlation between prototype aesthetics and color, pattern, shape, and high level. The simple feeling was negatively correlated with heavy presence. The awakening of past life experience was related to the understanding of creative intention and character creation. The semantic “conspicuous” (S3) was an important “bridge” to trigger other semantic evaluations.

Figure 6

AEN for each individual Yue Opera costume.

There was also a large degree of similarity in the centrality of clothing aesthetics of different characters (Figure 5b). For instance, the expected influence of archetypal aesthetic (A1) was the highest, and negative emotions (A2), heavy (S5), and simple (S2) had lower expected influence values. The network similarity index of the relational adjacency matrix and the centrality in Table 2 also reflect the generalizability of the conclusions to different Yue Opera costumes. The adjacency matrix correlation was between 0.62 and 0.72. The centrality correlation was even stronger, all above 0.90. In the AEN study [19], the adjacency matrix correlations of artworks ranged from 0.51 to 0.83, and the centrality matrix correlations ranged from 0.42 to 0.87.

However, the aesthetic network of Yue Opera costumes of different characters also had some differences. We then evaluated the variability of edge weights across the network to find the unstable relationships.

5.4 Variability network

Figure 7 represents a network where edges represent variability in connection strength between costumes. The edges below 0.05 were not shown, and the remained network was scaled from 0.05 to 0.11. In this network, the greatest variability raised between S4: high level and A4: knowledge (0.103), A1: archetypal aesthetic and S4: high level (0.095), and A5: interesting and S4: high level (0.095). As shown in Figure 6, the differences between these associations with high level (S4) and knowledge (A4) were mainly driven by the clothing of Laosheng, in the network of which the semantics “high level” (S4) and knowledge (A4) were positively correlated (0.274), and in the remaining four, they were weakly or hardly correlated (Huadan: 0.078; Shusheng: 0; Zhengdan: 0.103; and Wusheng: 0.059). The difference in associations between “high level” (S4) and archetypal aesthetic (A1) was mainly caused by the clothing of Laosheng (0.064) and Huadan (0.045). In these two networks, they were hardly correlated, and the remaining three networks had a positive correlation (Shusheng: 0.283; Wusheng: 0.133; and Zhengdan: 0.162). The difference in the associations between high level (S4) and interesting (A5) was mainly caused by the clothing of the Wusheng. The nodes “high level” (S4) and interesting (A5) included a positive association (0.219), while the other costumes were hardly associated.

Figure 7

Variability network.

Overall, these results suggest some differences in the associations between the aesthetic effects of works that rely on digital images of Yue Opera costumes, especially in semantic evaluation “high-level” and background knowledge understanding. Nevertheless, the basic structure of the AEN appeared to be relatively similar.

6.1 General network analysis

This study centered on the investigation of the aesthetic effects provided by simulated Yue Opera costumes portraying different roles. To achieve this, we applied the AEN model [19] to delve into the aesthetics of Yue Opera costumes. This approach allowed us to elucidate the network structure underlying the aesthetic impact of distinct role-based costumes. Notably, we identified a higher positive correlation between preferences (archetypal aesthetics) and color, pattern, and shape design elements, the semantic “high-level” as well as the cognition of aesthetic emotions (interesting). The influence of color on people’s perception and emotion had been discovered for a long time [60,61]. The finding about color, pattern, and semantics is partly in alignment with Han’s research [57]. These elements collectively formed the foundation of the aesthetic network inherent within Yue Opera costumes.

Of particular significance was the semantic descriptor “conspicuous,” which emerged as a pivotal “bridge” connecting the fundamental aesthetic network to the semantic evaluative network within Yue Opera costumes. The understanding of creative intent (A7), knowledge (A4), and positive emotion (A3) also played the role of “bridge” in the aesthetic network. Manipulating these nodes may change the relative ones in the whole aesthetic network. Therefore, it is significant to pay special attention to these contents in the process of costume design of Yue Opera.

In addition, a distinct interconnected group emerged, encompassing the comprehension of characterization, recollection of past experiences, and understanding of creative intent. A positive correlation was discerned among these elements. Their combined influence contributes significantly to the aesthetic network of Yue Opera costumes by facilitating the audience’s grasp of the contextual knowledge surrounding the costumes. This finding is somehow consistent with the conclusions drawn by Leder et al. [6]. The network configuration also resonates with the aesthetic triad pattern postulated by Chatterjee and Vartanian [33], which posited that the aesthetic experience is underpinned by a triadic framework encompassing the sensorimotor system, the emotion evaluation system, and the meaning knowledge system. Each of these systems possesses distinct neural foundations.

6.2 Role-based comparisons

In addition, we observed the difference in networks of role-based Yue Opera costumes. In the aesthetic network of Laosheng, high level (S4) and knowledge (A4) were positively correlated, while in the remaining four, they were weakly or not correlated. The possible explanation was that for Laosheng costumes, the audience understood the “high-level” sense of clothing from the top-down driven by cultural knowledge (the crane in the cloth is a symbol of longevity in Chinese culture), while other costumes may be more inclined to be viewed from the bottom-up based on the underlying design elements and other aspects. The difference in associations between high level (S4) and archetypal aesthetic (A1) was mainly caused by the clothing of Laosheng and Huadan. In these two networks, they were hardly correlated. This may indicate that the sense of “high-level” was not important in their archetype aesthetic evaluation. Specifically, in the aesthetic network of Laosheng, the archetype aesthetic was highly correlated with the underlying design elements (color, shape, and pattern) and interesting feelings. In the aesthetic network of Huadan, the archetypal aesthetic was closely related to the underlying design elements, positive emotions, and conspicuous feelings. The difference in the associations between high level (S4) and interesting (A5) was mainly caused by the clothing of the Wusheng. Through the later interview with the audience, we learned that the audience had a great interest in the dragon pattern in Wu Sheng’s clothing, which is the symbol of the emperor in Chinese history, making them have the association of “majesty” and “advanced” sense.

The establishment of this network model reminds us that to improve the dissemination and acceptance process of Yue Opera costumes, we can start from the design elements of costumes such as patterns, colors, and silhouettes to develop styles that audiences like [62,63]. It is also encouraged to strengthen the introduction of its background knowledge, arouse public’s past life experience, and convey the creator’s design intention. At the same time, we noticed that conspicuous (S3) semantic evaluation was an important node in the public’s aesthetic evaluation of such clothing. According to Chatterjee’s information processing model, attention is an important prerequisite for aesthetic evaluation [64]. Therefore, some details should be designed to be eye-catching in the digital costume design of Yue Opera.

6.3 Limitations and future directions

It should be noted that network analysis is a newly emerging research methodology in the field of psychological cognition in recent years. Inevitably, our research also possesses certain limitations. First, the relevant data were obtained from self-reports, which may have a certain social approval effect. A possible solution is to introduce objective measurement tools, such as eye trackers and electrocardiographs, to obtain a more realistic response from the audience. Second, the study was a cross-sectional design, and no causal conclusions can be drawn. Future research can examine the changes in the aesthetic structure through longitudinal tracking and can also test whether the intervention on aesthetic core emotions and cognition is more effective by improving the costume design of Yue Opera. Third, the main focus of this study is the youth group, and it is important to also consider the visual aesthetic preferences of both older individuals and the young to middle-aged groups toward Yue Opera. Moving forward, we aim to conduct age-stratified experiments to comparatively analyze the aesthetic preferences across various age groups. Fourth, in the aesthetics of Yue Opera costumes, accessories such as headpieces and shoes play a significant role. These accessories, including headpieces, possess unique design elements. Future research could incorporate Yue Opera’s headpieces, shoes, and other accessories into the study scope, making the research framework more comprehensive. Fifth, we designed five simulation images outfits for five common characters. However, in Yue Opera, the design elements of costumes can vary significantly among different characters. This factor was not considered in our current experiment. In future research, we plan to explore how changes in the design elements of a single character’s clothing might influence aesthetic experiences.