More knowledge causes a focused attention deployment pattern leading to lower creative performances

Previous studies demonstrate that people with less professional knowledge can achieve higher performance than those with more professional knowledge in creative activities. However, the factors related to this phenomenon remain unclear. Based on previous discussions in cognitive science, we hypothesised that people with different amounts of professional knowledge have varying attention deployment patterns, leading to different creative performances. To examine our hypothesis, we analysed two datasets collected from a web-based survey and a popular online shopping website, Amazon.com (United States). We found that during information processing, people with less professional knowledge tended to give their divided attention, which positively affected creative performances. Contrarily, people with more professional knowledge tended to give their concentrated attention, which had a negative effect. Our results shed light on the relation between the amount of professional knowledge and attention deployment patterns, thereby enabling a deeper understanding of the factors underlying the different creative performances of people with varying amounts of professional knowledge.

Supplementary Information Text S1. Statistical information of all variables in the regression models 1-5 Table S1 shows the statistical information on all variables that were used to build regression models 1-5 in the study.  Table S1 about here-----

S2. Details of the knowledge test and data collections of control variables in the survey data
In the survey data, we measured participants' amount of knowledge based on their scores on a speaker knowledge test. The full score of the knowledge test was 20. All 11 questions in the test were extracted from five well-known certification exams of speakers and audio devices in Japan. They were the: 1) Technical Skills Test of In the knowledge test, participants with higher scores were considered to have more professional knowledge of speakers; by contrast, participants with lower scores were considered to have less professional knowledge of speakers.
Moreover, to build the regression models in the study, we gathered participants' Previous research 1,2 has shown that along with control variables, the number of idea submissions should be controlled during the measurement of creative performances. Therefore, in this research, every participant was asked to submit only one idea during the survey. In other words, the number of idea submissions was controlled during the survey stage (i.e., the data collection stage) in this research.

S3.1. Verification of model-generated area ratios
To compute the area ratio in every picture accurately, we combined two effective methods in the computer-vision domain: Yolov3-Mobilenet and Grabcut [3][4][5] .
To verify these model-generated area ratios, we randomly selected 100 pictures from our datasets. The first author and a volunteer (who did not know the purpose of and was not involved in any analysis in this research) computed the area ratios of these 100 pictures by hand coding. The correlation between these two people's human-generated area ratios was 0.901 (p-value < 0.01), and the correlation between the model-and human-generated area ratios was 0.802 (p-value < 0.01).

S3.2. Verification of the relation between area ratios and attention deployment patterns
To verify the relation between area ratios and attention-deployment patterns, we implemented an additional analysis based on the survey data. Previous studies [6][7][8] ascertain that participants' attention-deployment patterns have a strong correlation with their ways of categorisation. When participants deploy their concentrated attention, they focus on the taxonomic relations during categorisation. By contrast, when participants deploy their divided attention, they focus on the thematic relations during categorisation. The taxonomic relations reflect the similarity of concepts based on categories, whereas the thematic relations are based on the same scenario or event 9 . For instance, dogs and bears are taxonomically similar because they belong to the same category (i.e., mammals); however, dogs and leads are thematically similar because they often occur in the same scenario (i.e., walking a dog).
Therefore, based on previous research 6 , we measured participants' ways of categorisation using 38-word-categorisation questions in the survey data, to verify the relation between area ratios and attention deployment patterns. In every question, participants were given a target word (e.g., dogs) and two words as choices (e.g., bears and leads) 6 . They were asked to select the one that they considered more similar to the target word. In every question, one choice was taxonomically similar to the target word while the other was thematically similar. By counting the number of taxonomically similar words chosen by participants, we could directly measure to what extent they focused on the taxonomic relations during their categorisation.

S4.1. Data collection of Wikipedia pages
To measure the idea novelty to history, we needed a dataset that generally includes speaker-related information in history. Based on previous research 10 , we utilised the data from Wikipedia. Using the same methods as in previous studies [11][12][13] , we defined the speaker-related pages by the citation relation in Wikipedia: We started from a seed-page, the page of the item of speaker in the Japanese Wikipedia (https://ja.wikipedia.org/wiki/%E3%82%B9%E3%83%94%E3%83%BC%E3%82%AB %E3%83%BC). We then collected all Wikipedia pages that were cited to explain the content on the seed-page (hereafter, referred to as the one-path pages). Next, we took these one-path pages as the new seed-pages and gathered pages that were cited by the one-path pages (hereafter, referred to as the two-path pages). We stopped at the twopath pages because previous studies [9][10][11] have found that information in the three-path pages is unrelated to seed-pages. Based on previous studies 9-11 , these Wikipedia pages comprised a sample of the speaker-related information in history.

S4.2. Metrics of creative performances based on the Amazon review data
In the Amazon review data, we built the metrics of creative performances based on review texts. Previous studies [14][15][16] have shown that the foundation of novel idea generation is a novel combination of information. We considered that the novelty of review texts reflected the ability of Amazon participants to find novel information combinations. Therefore, the review novelty was used as an indirect indicator of the creative performances of Amazon participants.
Since evaluating all 201,489 reviews is very time-consuming, previous studies 14,15 have shown that novelty to individuals is less important in creative performance evaluation; therefore, we only computed the review novelty to a group and review novelty in history based on the Amazon review data.
To build the metric of the review novelty to a group, we calculated the Tf-idf of every review text by comparing the words in the focused review text with those in all other review texts. Given the review collection R, a word w, and the focused review r, the Tf-idf of review r was calculated as follows: where , is equal to the number of times w appeared in r divided by the number of times all words appeared in r; | | is the number of all reviews; and , equals the number of reviews in which w appears. Since the Tf-idf of a word indicates how novel the word is compared with all words in other reviews, one review's novelty to a group (i.e., ) equals the sum of all words' Tf-idfs in the review 17,18 .
To build the metric of the review novelty in history, we gathered speaker-related Wikipedia pages in English Wikipedia according to the same method explained in Section S4.1. We finally gathered 16,863 speaker-related English Wikipedia pages.

S5. Different attention deployment patterns among people with different amounts of professional knowledge based on the Amazon review data
To examine the attention deployment patterns among people with different amounts of professional knowledge based on the Amazon review data, we first used a t-test to compare the area ratios between participants with distances to specialists at the bottom 25% (i.e., the high professional knowledge group) and those with distances to specialists at the top 25% (i.e., the low professional knowledge group). Because the area ratios ranged from 0 to 1, before the implementation of the t-test, we first transformed the area ratio using the arcsine transformation (i.e., = ( 2 )). As shown in Fig. S3, the high professional knowledge group had a significantly larger average area ratio than the low professional knowledge group (the average area ratios were 0.62 and 0.6 in the high and low professional knowledge groups, respectively; t = 1.67, p-value = 0.02).
Using the beta-regression model, we examined the relation between the distance to specialists and area ratio. In this model, we added the IDs of every target product as a dummy variable that only affected the constant. Therefore, the regression model estimated different constants for different target products (i.e., since there were 181 different types of speakers in our data, the model estimated 181 different constants).
In other words, in this regression model, we compared the different area ratios in the pictures of the same target product. In this way, we controlled the impact of the different target products on the area ratio. Additionally, previous studies 20, 21 have found that the length of the review text is closely related to participants' gender and age. Therefore, the number of words in the review was added as a control variable to indicate the potential difference between participants' gender and age. The statistical information and correlations among all variables are shown in Table S2 and Fig. S4.
As shown in Table S3, the regression results showed that the distance to  Table S3 about here-----

S6.1. Resampling of the idea novelty to individuals
Every idea under evaluation received one evaluation score from the seven evaluators.
We used the median of these evaluation scores to indicate the novelty of this idea to individuals 14,15 . Because the idea novelty to individuals had a skewed distribution with multiple peaks (as Fig. S5 shows), which would cause unreliable results in the statistical tests 22,23 , we implemented bootstrap resampling to reshape the distributions of the idea novelty to individuals. As shown in Fig. S5, given the high area-ratio group (i.e., participants with the top 25% area ratios) and the low area-ratio group (i.e., participants with the bottom 25% area ratios) under comparison, we resampled 60% (i.e., the resampling proportion; we also used 40% and 80% in the robust test) of the participants from the two groups and computed the means of their idea novelty to individuals in the bootstrap resampling. We repeated the resampling 100 times (i.e., the resampling times; we also used 200 times in the robust test) and, finally, obtained the resampled idea novelty to individuals of the two groups. As shown in the right panel in Fig. S5, the resampled idea novelty to individuals in both groups had an approximately normal distribution. Based on this resampled idea novelty to individuals, a statistical test could robustly reflect the difference in idea novelty between the two groups 16 . In Table S4, we show that the results in  Table S4 about here-----

S6.2. Impact of attention deployment patterns on creative performances based on the Amazon review data
To examine the impact of attention deployment patterns on creative performances based on the Amazon review data, we first analysed the average creative performance of participants with different area ratios. Fig. S6 shows the 1) average review novelty to a group and 2) average review novelty to history between the top area-ratio group (i.e., Amazon participants with the top 25% area ratios) and the low area-ratio group (i.e., those with the bottom 25% area ratios). The results were consistent with those in Fig. 3  In regression models S2-S3, we used the 1) review novelty to a group and 2) review novelty in history as dependent variables. The independent variable was the area ratio. As explained in the manuscript, since participants' gender, age, and amount of professional knowledge affect their creative performance, we added the distance to specialists and number of words in the review as control variables. The distance to specialists indicated each participant's amount of professional knowledge.
Research has shown 20,21 that the number of words in the review has a high correlation with a participant's gender and age; therefore, it was added to control the potential difference in a participant's gender and age. The statistical information on all variables in the regression models can be found in Table S2. The correlations and distributions among them are shown in Fig. S4. Since there was no significant collinearity among the independent and control variables, we used the linear regression (OLS) to build the models. The results of the regressions are shown in Table S5. We found that the area ratio had significantly negative effects on creative performance (in the model using the review novelty to a group as the dependent  Table S6 (based on the survey data) and Table S7 (based on the Amazon review data). Results showed that the amount of professional knowledge did not have a significant direct impact on creative performance. However, by affecting the attention deployment pattern, the amount of professional knowledge negatively impacted creative performance.
Therefore, the above results indicate that participants with more (less) professional knowledge deployed their concentrated (divided) attention that led to their lower (higher) creative performance.

S8. The consistency of evaluations among experts
To measure the idea novelty of individuals, we asked seven experts of speakers to evaluate participants' ideas. To justify the robustness of these evaluations, we investigated the correlations among the seven experts' evaluation scores. As Fig. S7 shows, significantly positive correlations, ranging from 0.26 to 0.63, were found among the experts' evaluation scores. These results indicate that for the same idea, the seven experts provided consistent evaluations on its novelty. Fig. S1. Distribution, scatter plots, and correlations of all variables in regression models 1-5. The names of the variables are all to the left of the figure. In this figure, the diagonal shows the histograms, with the density curve of every variable. Graphs in the lower triangle show the scatter plots between each pair of variables. The red line shows the relation between these two variables predicted by the linear regression (OLS). The ellipses show the correlation ellipses, with the centre of each as a red point. In the upper triangle, the correlations between each pair of variables are shown. One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01.   One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01.     Table S3. Regression results of the area ratio based on the Amazon review data.

Dependent variable
Area ratio Model S1 Distance to specialists −0.961 *** (0.189) Number of words in the review (in logarithm scale) 0.083 *** (0.016) phi 1 1.0 *** (0.02) Average constant 2 1.354 Observations 4,857 R-square 0.007 Log Likelihood 3,644.402 Note: One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01; parentheses indicate the standard error of every variable.  Note: One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01; parentheses indicate the standard error of every variable. One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01; parentheses indicate the standard error of every variable; the standard error of every coefficient is shown in '().' Table S7. Results of path analysis based on the Amazon review data.

Attention deployment pattern-Professional knowledge Dependent variable
Area Ratio Distance to specialists -0.029 ** (0.014) Creative performance-Attention deployment pattern + Professional knowledge Dependent variable Review novelty to a group Review novelty in history Area ratio -0.193 *** (0.014) -0.135 *** (0.014) Distance to specialists 0.002 (0.014) -0.004 (0.014) Observations 4,857 Note: One asterisk refers to a p-value smaller than 0.1, two asterisks to a p-value smaller than 0.05, and three asterisks to a p-value smaller than 0.01; parentheses indicate the standard error of every variable; the standard error of every coefficient is shown in '().'