https://orcid.org/0000-0002-9309-2834
Figures
Abstract
Population-level national networks on social media are precious and essential for network science and behavioural science. This study collected a population-level Twitter network, based on both language and geolocation tags. We proposed a set of validation approaches to evaluate the validity of our datasets. Finally, we re-examined classical network and communication propositions (e.g., 80/20 rule, six degrees of separation) on the national network. Our dataset and strategy would flourish the data collection pool of population-level social networks and further develop the research of network analysis in digital media environment.
Citation: Guan L, Liu XF, Sun W, Liang H, Zhu JJH (2022) Census of Twitter users: Scraping and describing the national network of South Korea. PLoS ONE 17(11): e0277549. https://doi.org/10.1371/journal.pone.0277549
Editor: Rashid Mehmood, King Abdulaziz University, SAUDI ARABIA
Received: January 4, 2022; Accepted: October 30, 2022; Published: November 17, 2022
Copyright: © 2022 Guan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data are available from the Harvard Dataverse (DOI: doi.org/10.7910/DVN/9GRCYU; URL: https://doi.org/10.7910/DVN/9GRCYU).
Funding: This study was supported by Shanghai New Media Center at Fudan University in the form of a grant awarded to LG. This study was also supported by Shanghai Pujiang Program in the form of a grant (22PJC012) awarded to LG. This study was also supported by Centre for Communication Research at City University of Hong Kong in the form of a grant (9360120) awarded to JZ. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Social network analysis (SNA) has been fuelling social sciences for decades. Through the use of graph theory and large network analytical techniques, the research foci of SNA differ from traditional social science regimes, shifting from social actors (e.g., individuals and organizations) to their social relationships and surrounding environments [1].
Although SNA has been employed in social media studies to examine online network patterns in the past decade, national-level population-based online social networks are still rare to find but of great importance to network studies. First, current online population networks on social media are often found to be either arbitrarily or narrowly defined, e.g., users discussing the same topics, followers of a social media account, and social media accounts of the members of an organization [2–4], whereas population-level social networks at more stable and general levels are rare to find. The causal relations and structural patterns of specific topic or event networks could hardly extend to the situations of other topics, events, or organizations. The findings relating to different topics or events are often found to result in different and even contradictory conclusions. Thus, current SNA studies require large-scale social networks at more stable community levels (e.g., nation, ethnic groups) to observe structural patterns of higher generalizability.
Second, large-scale population-based networks require precise validation methods to guarantee the completeness of datasets. Data collection of population-based networks needs to cover every actor and all their relationships in the entire population. It requires a clear boundary of the population belongingness [5]. When adopted in offline studies, the boundary is often defined according to geolocation or social identity attributes, such as students of a university or residents of a village [6]. When adopted in online studies, this network boundary becomes rather difficult due to the lack of a sampling frame for the population structure of social media users [7]. Thus, validation solutions still need to sought to ensure the data completeness.
Third, a batch of network structure propositions and social theories remain to be examined and uncovered in the national-level population social networks. Social science scholars used to rely on the national census datasets conducted by the government to capture an accurate and comprehensive overview of people’s offline lifestyle [8]. However, when depicting the online lifestyle, sampled data often become the only choice, due to the lack of sampling frames for the online population structure and the technical constrictions on data collection and storage [9]. As a result, a great number of network propositions and social science theories have few chances to be examined and summarized at the national-level social networks.
The novelty and originality of the present study lies in that, in order to enrich the national online network pools, we collected a full national social media user network, South Korea, on Twitter and proposed two validation methods to help evaluate the validity of our data collection strategy. Furthermore, to fill the research gap that online social theories were mostly examined on sampled data rather than full population network datasets, we re-examined classical network and communication propositions (e.g., 80/20 rule, six degrees of separation, etc.) on the national social network. The present study could provide practical implications by serving as a reference to help validate existing network sampling methods. Besides, the study could also bring theoretical insights by extending the classical theories contexts to online social environment.
Related work
Previous literature on national-level Twitter accounts mainly follows two streams of data collection strategy: location-based and language-based. Bruns, Moon, Münch, & Sadkowsky scrapped the first national Twitter network in 2017 [10]. Benefiting from the comparative distinctiveness of Australian timezones and user profile geolocation tags, they successfully depicted the general network patterns of Australian “Twittersphere”. However, later, Twitter conducted a series of data protection changes to their developer APIs. Those properties, including geotags, user time zone, and the interface language used on Twitter, are all inaccessible by now [11].
The later national Twitter networks mainly utilized the language-based data collection strategy. Bruns & Enli collected the entire national Twittersphere of Norway based on the Norwegian language of user profile [12]. Bruns and others scraped the German Twitter network dataset TrISMA [13] and later the dataset served as a ground truth for the so-called rank degree sampling method [11]. Recently, Munch and Rossi compared two national Twitter networks, Italian and German, based on the language detection of user profile [14]. Our data collection mainly followed the language-based approach. Meanwhile, we also examined the geolocation information that user self-reported in their profiles. Besides, our work proposed new validations methods that could largely help validate the results.
Materials and methods
Ethics statements
The study was approved by the Human Ethical Review in College of Liberal Arts and Social Sciences, City University of Hong Kong. All the data collected in this study are open for the public. Data was obtained from Twitter’s REST API. Developer account was granted by Twitter before data collection started, which allows the access to the data. Identifier data fields were replaced by unidentifiable pseudo code after all data are collected upon the end of the project.
Data collection
This current study evidenced a data collection strategy to scrape a national social media network on Twitter. The reasons we choose Twitter as the example platform are threefold. Firstly, Twitter could serve as a typical social media example, as it combines the fundamental functions of social media: both social networking services and information sharing application [15–17]. Users can both make friends with other users by following their accounts and send out their own tweets. Besides, Twitter is one of the most popular social media around the world, with 330 million monthly active users and 500 million tweets sent out per day [18]. Finally, Twitter provides advanced APIs to help researchers obtain the network data.
The reasons why we select the South Korea as the case of the nation are based on two aspects of considerations: the uniqueness of its official language and the population size of the country. Firstly, South Korea is the only country on Twitter that utilizes the Korean language as the official language. The uniqueness of its language could help us to identify the nationality of the users. Secondly, the population of South Korea was estimated at 51 million, ranking the 28th in the world [19]. Compared with other countries that fit the language requirement (e.g., Japan), the population size of South Korea is more feasible for the data collection procedure.
We used Twitter’s REST APIs to collect the profiles, following relationships, and timelines of the Twitter users. The data collection of the socio-centric network requires a clear boundary to help identify the nodes within the network. In this study, we utilized the language and geolocation information that users filled in their twitter profiles to help identify the nationals of South Korean. Specifically, if the user utilized Korean as his/her language on Twitter or had filled his/her geolocation tag as South Korea or any city in South Korea, s/he will be labeled as a South Korean Twitter user. Besides, we also categorized the users according to their activeness on Twitter. Users who had updated their last posts in 2019 will be identified as active Twitter users, whereas inactive users will be excluded from our sample.
We utilized the snowball sampling method to collect the population of active South Korean Twitter users. Fig 1 illustrates the specific data collection procedures. In the first step, a list of the most popular Twitter accounts in South Korea was obtained from the Socialbakers’ Twitter statistics [20]. From this popularity list, 28 popular accounts in the categories of brand, society, and sports were selected as the seeds of the snowball sampling. Then, in the second step, the profiles of these popular accounts’ followers were collected through Twitter APIs. If the follower was identified as an active South Korean user in 2019, s/he will be kept as ego users in the first batch.
In the third step, we obtained the first batch ego user’s following relationship lists and tweets/ retweets in user timelines. If the alters (followers/ followees) of first batch ego users were identified as active South Korean users, they were kept and labeled as second batch ego users. By continuously repeating step 2 and step 3, we obtained profile, following relationship, and timeline data of active South Korean users until batch N.
Table 1 shows the numbers of new Twitter users, new South Korean users, and new active South Korean users scraped for each batch. After conducting six batches of snowball scraping, the increase rate of new South Korean users has declined to 1.26%. Up until the sixth batch, we obtained 127 million Twitter users’ profiles, with 18.1 million of them identified as South Korean users and 2.59 million of them as active South Korean users in 2019.
To discuss the threshold for suspending the snowball scraping, we further conducted another 20 batches after the sixth batch. Fig 2 shows the accumulated number of South Korean Twitter users after each batch of snowball sampling scraping. After the third batch, the increasing trend gradually slowed down, and then the number began presenting a stable trend after the fifth batch. The increase rates after the sixth batch are all less than one percent, which is small enough to suspend the snowball scraping procedure.
Data validation
We proposed two approaches to help evaluate the validity of our dataset. The first one is to roughly compare with public statistics of Twitter usage in South Korean. Considering the population of South Korea was estimated at 51.2 million people in 2019 [19] and, the penetration rate of Twitter was 31% in 2020 for South Korean [21]. Thus, there should be at least 15.872 million Twitter users in South Korea. Previous studies on Twitter bots stated that up to 15% of Twitter accounts are bots, not people [22]. Thus, at least 18.2 million South Korean accounts (including the South Korean bot accounts) were expected to be scrapped. The figure in our results (18.3 million South Korean users by the twenty-sixth batch) was slightly larger than that we deducted from the statistics.
Then, to further validate our datasets, we devised a tweet matching validation approach to help validate the completeness of the tweet datasets. The basic idea of this approach is to compare the number of tweets that include a certain hashtag or keyword within our collected tweet dataset with the results that we could obtain from the Twitter firehose API by searching the same hashtag or keyword. For example, in our case, we aim to compare Korean language hashtag tweets in our collected timeline dataset with the Twitter firehose API results of the same hashtag. Twitter firehose API guarantees full access to the tweets that match the search criteria. Searching certain keywords on the Twitter firehose API and restricting the period to 2019 would provide 100% of the tweets containing the keywords in 2019. Meanwhile, our data collection strategy aims to reach the whole population of South Korean users who had tweeted in 2019. Thus, theoretically speaking, every tweet in the Korean language in 2019 should be found in our tweet results.
The specific validation procedure included three steps. In the first step, we extracted a list of popular Korean language hashtags from the tweet timeline dataset that we had scraped. In the second step, we collected all the tweets containing the popular Korean hashtags in 2019 from a commercial data analytics platform, Crimson Hexagon. The platform could get access to the Twitter firehose API and provide all the tweets by keyword searching. Then, in the third step, we examined the differences between the hashtag search results from Crimson Hexagon and the timeline tweet data in our network sample.
First, we identified the top 5,000 popular Korean language hashtags from our datasets. Surprisingly, we found that most of these hashtags were related to K-pop celebrities, K-pop events, and K-pop culture products. As K-pop culture is extremely popular and influential across the world [23], hashtags relating to K-pop culture could be heavily contributed by both South Korean fans and fans from other countries. Thus, we manually excluded K-pop related hashtags and identified 88 Korean language hashtags that were related to South Korean social events or everyday life, such as religion, food, domestic location, and domestic events. We expected that these hashtags were mainly contributed by South Korean users. The hashtags and English translations by Google Translation appear in Table 2.
In the second step, we collected all the tweets covering the Korean hashtags via the Crimson Hexagon platform. As shown in Table 2, the column labeled “Number of tweets downloaded via CH” represents the hashtag searching results via the Twitter Firehose API with the language restriction as Korean, and the column labeled “Number of tweets included in our dataset” represents the results from our timeline dataset. Of note, the Twitter Firehose API could provide both a summary number of hashtags and access to download all the related tweets. The total summary number of hashtags covers not only existing tweets but also tweets that have been deleted by users. However, these deleted tweets could not be accessed and downloaded through the Twitter REST APIs that we used for the data collection. Thus, to compare the results efficiently, the number of tweets via Crimson Hexagon shown in Table 2 represents the number of existing tweets that could be downloaded at the end of 2019.
Finally, we compared the Twitter Firehose API results with our datasets. Importantly, when scraping users’ timeline data, Twitter REST APIs only allow developers to download the latest 3,200 tweets for one user. Thus, we could not guarantee that we had downloaded all the timelines of users in 2019 in our dataset. The solution to this problem was that we turned to examine whether the author of the tweets from the Twitter Firehose API results was included in our 2019 active South Korean user pool. If the author was covered by our data collection procedure via 26 batches of the snowball sampling approach, then the tweets were identified as included in our dataset. The results illustrated in Table 2 show that the coverage rates of our dataset range from 51.04% to 97.97%, with an average score of 84.4%.
We further looked into the tweets that were not covered by our datasets. Totally there are 165,989 tweets that were not covered by our data collection, and they are contributed by 20,608 Twitter accounts. Fig 3 shows the distribution of reasons why our datasets did not cover these accounts. We judged these accounts in two dimensions: whether these accounts had been covered in our snowball sampling procedure and whether these accounts showed any hints of nationality in their Twitter profiles. According to the results, 69.8% (14,380) of these accounts had neither been included in our sampling procedure nor showed any traces of South Korea in their profile information. Although these accounts had posted one or more tweets with Korean language hashtags, we could not guarantee the nationalities of these users. Since they neither wrote their Twitter names and descriptions in Korean characters nor set their locations within South Korea, they could possibly be foreigners tweeting on certain Korean issues. Furthermore, 10% (2,068) of the identified accounts were found to have been included in our snowball sampling procedure. However, during our data collection and nationality identification procedure, these accounts were labeled as not South Korean users due to a lack of South Korean traces in their profile information. Similar to the previous category of accounts, we could not guarantee their nationality since there was a chance that these accounts could be foreigners who were interested in Korean issues. Finally, there were 4,160 (20.2%) active South Korean users that should have been included in our datasets but were unfortunately not. Possible reasons could be that these accounts were relatively isolated in the South Korean user network. These accounts neither followed the most popular South Korean Twitter accounts from which we started our snowball sampling nor had any South Korean friends on Twitter (we would have covered them if they had any domestic friends). Given that we aimed to explore national patterns of the connected network structures, these isolated accounts should have minimally affected our results.
Results
Network structure
Power-law distribution in follower-followee network.
We constructed the active South Korean Twitter user network in which nodes were active South Korean Twitter users, and edges were their following relationships. There were 2,309,959 nodes and 164,476,927 directed edges within the active South Korean Twitter network. Fig 4 displays the distribution of the number of followers and friends. The y-axis represents the complementary cumulative distribution function. Similar to previous literature [24, 25], our findings observe that both the number of followers (in-degree) and the number of friends (out-degree) on Twitter are unequally distributed with heavy tails, resembling power-law distributions (αfollower = 2.09; αfriend = 2.09).
Network attributes of the follower-followee network.
In addition to the power-law distribution of follower-followee networks, some network attributes are frequently measured to depict the basic tie formation in the population-based following network in Twitter. Reciprocity, assortative mixing, and transitivity were calculated and compared with previous results of Twitter networks.
Reciprocity quantifies the likelihood of nodes to be mutually linked [26]. As reciprocity captures the symmetric form of interactions between two users, scholars believe the overall score of reciprocity could indicate how likely users take advantage of the platform’s social networking function and construct mutual interactions through the platform [27]. An early study in 2010 found that Twitter had a relatively low reciprocity of 22.1% [24], compared with other social media platforms with networking service (e.g., 68% on Flickr and 84% on Yahoo! 360) [28, 29]. The finding of low reciprocity for Twitter, as well as other findings on follower distribution and effective diameter, indicate that Twitter is more like news media rather than social media. Users on Twitter pay more attention to the news-related functions rather than constructing mutual friendships with others. However, a later study in 2014 overturned the previous conclusion that they found active user networks on Twitter actually had a much higher reciprocity level of 42% [30]. Our national network results find the reciprocity of the active South Korean network in our dataset was 37.86%. Compared to existing findings, the national Twitter network shows a moderate level of reciprocity, which indicates a moderate level of Twitter’s social function usage in the national network.
Assortative mixing indicates the preference for nodes in a network to link with other similar nodes [30]. Degree assortativity was regarded as a fundamental attribute to distinguish the social networks from other types of large-scale networks [31], thus it is one of the essential measures to help identify the social function of a network. In most undirected social networks, a typical degree assortativity ranges between 0.1 to 0.4. Previous literature found the directed networks (e.g., Twitter, Google+, etc.) have relatively low assortative scores. Existing reports of assortativity for Twitter range from neutral to quite disassortative (-0.053 to -0.88) [32]. The assortative measure of our national network is -0.0455, indicating a relatively moderate level of node homophily in the South Korean Twitter network.
Transitivity (also called clustering coefficient) measures the extent to which nodes in the network tend to cluster together and form closed triplet relationships. Research on social dynamics pays specific attention to the formation of friend-of-a-friend relationship [33], because the overall transitivity score reveals the existence of tightly connected communities (clusters, subgroups, cliques, etc.) and may influence the stability and development of the network in the long term [34]. We measure the transitive clustering coefficient for the directed network of Twitter. The score (tcc = 2.58%) is a bit higher than previous findings (1.9% in [35]). This indicates that the active Twitter user network of South Korea are more likely to cluster together and form connected triangle relationships.
Six degrees of separation.
The notion that people are connected through “six degrees of separation” is often used as a synonym for the idea of the “small world” phenomenon [36]. Degree of separation has been re-examined for many online networks, for example, e-mail networks [37], instant message networks [38], and webpage networks [39]. Previous literature suggests that the average path length of online social networks ranges from 3.43 to 4.47 [40–42]. In particular, users on Twitter are connected with an average degree of separation of 4.12 [24]. Our findings are similar to previous studies. Fig 5 shows that the average path length is 4.05, and 98.8% of user pair distances fall in five steps. The average path length is quite short considering the size of national Twitter network. That indicates the national network of Twitter users has a relatively close connectedness and exhibits some small-world properties.
Usage
80/20 rule of content generation.
The 80/20 rule, also called the Pareto principle, originally refers to the phenomenon that approximately 80% of the land in Italy belonged to 20% of the population [43]. The rule was later generalized to both ownership, productivity, and other behaviour patterns in previous research on business, epidemic, and social media. For example, a small proportion of an organization’s marketing units often generates a large proportion of the profits [44]; 80% of epidemic transmission is accounted for by 20% of individuals [45]; more like a 76/20 proportion in UGC production on Internet [46], etc. Concerning the productivity on social media in particular, previous literature observed a power-law distribution with an exponent of about 1.92 on Twitter [47].
Our results could not replicate the power-law findings on Twitter. Fig 6 illustrates that the distribution of the number of tweets per user follows a resemble exponential distribution than a standard power-law distribution. However, we did observe an approximate 80/20 rule based on Fig 7. About 20% of users created 92% of tweets, or 80% of contents are posted by 10% of users. More than half of the users do not make any obvious contributions to online content production.
Originality, sociability, and syntactic use.
In previous literature, several syntax records and interaction features are frequently mentioned to describe how users post tweets: retweet, reply, hashtag, and URLs. They are commonly related to the research focus of information originality [48], sociability [49], and syntactic usage [50]. First, South Korean ranks in the forefront of retweet rate among multiple countries or cultural communities. The proportion of retweets in Korean language is 78% in 2018, lower than Thai (96%), but significantly higher than English (40%) and the representative random Twitter dataset (22.4%) [25, 51]. In our dataset, the percentage is 58%, which indicates a relatively low originality rate of South Korean Twittersphere compared to other cultures’.
Reply and @mention functions have been commonly regarded as indicators for conversations in Twitter [25]. The estimated proportion of replies and mentions in Korean-language tweets (ranging from 40% to 59% for replies and 73% for mentions) were relatively higher than other language, e.g., English (29% for replies and 47% for mentions) and all language (24% for replies and 49% for mentions) [25, 52]. In our datasets, 15.4% tweets are replies and 68.9% tweets contains @-mentions. As @-mentions could be caused by both retweet and reply-to. After excluding these, 2.1% tweets mentioned other users’ ID.
In terms of syntactic usage, previous research found there is a relatively moderate level of URLs (17%) and hashtag (11%) use in Korean language tweets, lesser than English language tweets (25% for URLs and 14% for hashtags), but similar to all language tweets (21% for URLs and 11% for hashtags) [52]. Our findings suggest that 14.7% Korean language tweets contain URLs and 23.8% tweets contain hashtags.
Overall, our findings indicate that there is a low-level originality (excluding retweets and replies: 26.3%) and a low-level sociability (reply+@mention: 17.5%) for South Korean users on Twitter. Twitter serves more like an information sharing site rather than social interactional platform for South Korean community.
Circadian rhythms.
Consistent with previous findings on digital device and social media usage, Twitter message activity was found to follow the daily and weekly circadian rhythm [53]. Twitter usage starts to rise in the morning, continue to increase during the daytime, and peak in the evening [25]. Besides, previous research observed different usage patterns between weekdays and weekends for individuals with different demographic characteristics [54]. Our results are consistent with previous studies on daily circadian rhythm. Fig 8 illustrates that Twitter messaging activity in South Korean network reaches the peak around 9:00 pm for both weekdays and weekends. Besides, we also observe that Twitter usage at weekends were lower in the morning, but higher at noon and night, compared to weekday activities.
Topics & social bots.
To better understand the users’ profile and their topic attention, we further looked into their timelines. We conducted the unsupervised topic modelling approach on a random sample of 10,000 users’ timeline (more than 340k tweets). To better interpret the results, we firstly translated the tweets into English and applied the state-of-art topic modelling technique, BERTopic [55], to classify the topic. By manually interpreting the meaning of each topic, we classified all the tweets into 6 categories. Among these tweets, 30.1% were recognized as related to K-POP and celebrity activities, followed by fashion and art (23.3%), technology (16.5%), daily life (13.9%), game & sport (8.4%), and other uncategorized personal statements (7.8%).
Besides, we detected the average Twitter bot rate in active 2019 South Korean datasets using a 30,000 random user sample. Using a widely-used library BotometerLite [56], we found that 51.5% of accounts were identified as bots with a botlikelihood score of .5 or higher. The corresponding rate changes to 42.1% for a threshold of .6, 33.0% for a threshold of .7, and 21.7% for a threshold of .8. The bot rate is much higher than previous results for English-language tweets (13.2% [57] and 15% [22]). However, considering the culture difference of South Korean [58–60], our results are quite similar to previous results for Asian-language tweets (52% for Indonesia, 52% for Philipplines, 71% for Singapore, and 63% for Taiwan, with a BotHunter threshold of .5) [61].
Conclusions and discussions
The current study evidenced a data collection strategy to scrap and validate a national-level Twitter network. To collect the population of active South Korean Twitter users, we conducted twenty-six batches of data collection by utilizing the snowball sampling method. Then, we proposed two approaches to help evaluate the validity of our dataset. The first one is to roughly compare with public statistics of Twitter usage in South Korean. Then, to further validate our datasets, we devised a tweet matching validation approach to help validate the completeness of the tweet datasets. After comparing the number of tweets that include a certain hashtag or keyword within our collected tweet dataset with the results that we could obtain from the Twitter firehose API by searching the same hashtag or keyword, we found a relatively good coverage rates of our dataset, with an average score of 84.4%. Finally, we re-examined a series of network patterns and social theories in the full social network of South Korean, including network attributes (network distribution, reciprocity, assortative mixing, transitivity, degree of separation, etc.) and usage patterns (originality, sociability, syntactic use, circadian rhythms, topics distribution, social bots, etc).
The study could bring practical contributions to network analysis research in two ways. First, this study evidenced a feasible and flexible strategy to scrap the national Twitter networks and devises two validation approaches to verify the completeness of the collected network. Although we only demonstrate one case, it could be flexibly applied to many other nations or ethnic groups. For example, the proposed approach could be directly utilized to other nations that use their own unique language as the mainstream communication language, e.g., Japan, Bulgaria, etc. The strategy would largely flourish the data collection pool of population-level social networks and further develop the research of network analysis in digital media environment.
Second, national online social networks could serve as a reference to help validate existing network sampling methods. Probability sampling in network science are difficult due to the lack of a sampling frame for the population structure of social media users [7]. Many prevailing sampling methods (e.g., random walks) were proposed to conduct ego selection, where the data collection starts with some randomly selected seed nodes and continues with the random selection of links from the seeds [62]. However, some of these methods (e.g., the initial version of the random walk) have been questioned to be systematically biased toward the hub nodes, as it cannot solve the self-selection bias inherent in the link structures of a network [63, 64]. In this case, large-scale population-based networks are in crucial needs to help validate the network sampling biases. The national social network is one of the optimal labelled reference for network sampling validation, as it is the largest community-level networks with the basis of a common territory, language, and culture.
Finally, current study could also provide theoretical implications to existing social science literature. We re-examined classical network patterns and theories on the national social network of South Korea and found new patterns on online full population network. For example, our works show that the national network of South Korea has a relatively low degree of separations, which indicates a close connectedness and “small-world” properties for the South Korean society. Future studies could further examine the connectedness of other population-level social networks and estimate how the extents of “small-world” effects vary among different societies.
Limitations and future studies
Our study is subject to some limitations. First, we note that our dataset could still contain false positives and omit false negative: it would also match false-positive accounts such as foreigners who lived in South Korea. In addition, there may be South Korean users never using Korean language on Twitter. We admit this existing situation and this is the common problem for all so-called national-level Twitter networks in previous literature. This study have utilized two validation methods to help minimize it.
Another obvious limitation is about the language challenge. As non-Korean researchers, we employed English translations of sampled tweets in Korean to conduct the topic modeling. The results might be more correct if applying BERT to tweets in Korean without first translation. Future studies could apply native language packages to help validate our results.
Besides, in this study, we only considered a single social platform for collecting the population-level networks. Future directions could replicate or further develop the approach to other social networking sites. Finally, current study only examines some descriptive patterns and findings on the population-level social network. Future studies would continue to investigate on other correlational or causal patterns of the population network, e.g., the causes of population networks formation, the relationship between network structures and functionalities, etc.
References
- 1. Zhang M. Social network analysis: History, concepts, and research. Handbook of social network technologies and applications: Springer; 2010. p. 3–21.
- 2. Bailo F, Vromen A. Hybrid social and news media protest events: from# MarchinMarch to# BusttheBudget in Australia. Information, Communication & Society. 2017;20(11):1660–79.
- 3. Huang R, Sun X. Weibo network, information diffusion and implications for collective action in China. Information, Communication & Society. 2014;17(1):86–104.
- 4. Rodrigues UM, Niemann M. Social media as a platform for incessant political communication: a case study of Modi’s “clean India” campaign. International Journal of Communication. 2017;11:23.
- 5. Perkins JM, Subramanian S, Christakis NA. Social networks and health: a systematic review of sociocentric network studies in low-and middle-income countries. Social science & medicine. 2015;125:60–78. pmid:25442969
- 6. Himelboim I. Social network analysis (social media). The international encyclopedia of communication research methods. 2017:1–15.
- 7. Mouw T, Verdery AM. Network sampling with memory: a proposal for more efficient sampling from social networks. Sociological methodology. 2012;42(1):206–56. pmid:24159246
- 8. Tu W-J, Zeng X, Liu Q. Aging tsunami coming: the main finding from China’s seventh national population census. Aging clinical and experimental research. 2021:1–5.
- 9. Lehdonvirta V, Oksanen A, Räsänen P, Blank G. Social media, web, and panel surveys: using non‐probability samples in social and policy research. Policy & internet. 2021;13(1):134–55.
- 10. Bruns A, Moon B, Munch F, Sadkowsky T. The Australian Twittersphere in 2016: Mapping the Follower/Followee Network. Social Media + Society. 2017;3(4). WOS:000443463700015.
- 11. Munch FV, Thies B, Puschmann C, Bruns A. Walking Through Twitter: Sampling a Language-Based Follow Network of Influential Twitter Accounts. Social Media + Society. 2021;7(1). WOS:000610521600001.
- 12. Bruns A, Enli G. The Norwegian Twittersphere Structure and Dynamics. Nordicom Review. 2018;39(1):129–48. WOS:000433247600009.
- 13. Axel Bruns JB, Banks John, Tjondronegoro Dian, Dreiling Alexander, Hartley John, Leaver Tama, Aly Anne, et al. TrISMA: Tracking infrastructure for social media analysis. QUT Digital Media Research Centre; 2016.
- 14. Münch FV, Rossi L. A TALE OF TWO TWITTERS? IDENTIFYING BRIDGES BETWEEN LANGUAGE BASED TWITTERSPHERES. AoIR Selected Papers of Internet Research. 2020;2020(0).
- 15. Bruns A, Stieglitz S. Towards more systematic Twitter analysis: metrics for tweeting activities. International journal of social research methodology. 2013;16(2):91–108.
- 16. Saura JR, Palacios-Marqués D, Ribeiro-Soriano D. Using data mining techniques to explore security issues in smart living environments in Twitter. Computer Communications. 2021;179:285–95.
- 17. Saura JR, Palacios-Marqués D, Ribeiro-Soriano D. Exploring the boundaries of open innovation: Evidence from social media mining. Technovation. 2022:102447.
- 18.
Omnicore. Twitter by the numbers: Stats, demographics & fun facts 2019. Available from: https://www.omnicoreagency.com/twitter-statistics/.
- 19.
Worldometers. South Korea population 2019. Available from: https://www.worldometers.info/world-population/south-korea-population/.
- 20.
Socialbakers. Twitter statistics for South Korea. 2019. Available from: https://www.socialbakers.com/statistics/twitter/profiles/south-korea.
- 21.
Statista. Penetration of leading social networks in South Korea as of 3rd quarter 2020 2020. Available from: https://www.statista.com/statistics/284473/south-korea-social-network-penetration/.
- 22.
Varol O, Ferrara E, Davis C, Menczer F, Flammini A, editors. Online human-bot interactions: Detection, estimation, and characterization. Proceedings of the international AAAI conference on web and social media; 2017.
- 23. Kim M, Heo Y-C, Choi S-C, Park HW. Comparative trends in global communication networks of# Kpop tweets. Quality & Quantity. 2014;48(5):2687–702.
- 24.
Kwak H, Lee C, Park H, Moon S, editors. What is Twitter, a social network or a news media? Proceedings of the 19th international conference on World wide web; 2010.
- 25. Liang H, Fu K-w. Testing propositions derived from Twitter studies: Generalization and replication in computational social science. PloS one. 2015;10(8):e0134270. pmid:26287530
- 26. Floría LM, Gracia-Lázaro C, Gómez-Gardenes J, Moreno Y. Social network reciprocity as a phase transition in evolutionary cooperation. Physical Review E. 2009;79(2):026106. pmid:19391805
- 27.
Holton AE, Baek KH, Coddington M, Yaschur C, editors. Soliciting Reciprocity: Socializing, Communality, and Other Motivations for Linking on Twitter. International Symposium on Online Journalism, Austin, TX, April; 2013.
- 28.
Cha M, Mislove A, Gummadi KP, editors. A measurement-driven analysis of information propagation in the flickr social network. Proceedings of the 18th international conference on World wide web; 2009.
- 29. Kumar R, Novak J, Tomkins A. Structure and evolution of online social networks. Link mining: models, algorithms, and applications: Springer; 2010. p. 337–57.
- 30.
Myers SA, Sharma A, Gupta P, Lin J, editors. Information network or social network? The structure of the Twitter follow graph. Proceedings of the 23rd International Conference on World Wide Web; 2014.
- 31. Newman ME, Park J. Why social networks are different from other types of networks. Physical review E. 2003;68(3):036122.
- 32. Bhattacharya S, Sinha S, Roy S. Impact of structural properties on network structure for online social networks. Procedia Computer Science. 2020;167:1200–9.
- 33. Klimek P, Thurner S. Triadic closure dynamics drives scaling laws in social multiplex networks. New Journal of Physics. 2013;15(6):063008.
- 34. Jin EM, Girvan M, Newman ME. Structure of growing social networks. Physical review E. 2001;64(4):046132. pmid:11690115
- 35.
Trolliet T, Cohen N, Giroire F, Hogie L, Pérennes S, editors. Interest clustering coefficient: a new metric for directed networks like twitter. International Conference on Complex Networks and Their Applications; 2020: Springer.
- 36. Uzzi B, Amaral LA, Reed‐Tsochas F. Small‐world networks and management science research: A review. European Management Review. 2007;4(2):77–91.
- 37. Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’networks. nature. 1998;393(6684):440–2. pmid:9623998
- 38.
Leskovec J, Horvitz E, editors. Planetary-scale views on a large instant-messaging network. Proceedings of the 17th international conference on World Wide Web; 2008.
- 39. Barabási A-L. Network science. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences. 2013;371(1987):20120375. pmid:23419844
- 40. Edunov S, Diuk C, Filiz IO, Bhagat S, Burke M. Three and a half degrees of separation. Research at Facebook. 2016;694.
- 41.
Backstrom L, Boldi P, Rosa M, Ugander J, Vigna S, editors. Four degrees of separation. Proceedings of the 4th Annual ACM Web Science Conference; 2012.
- 42.
Bakhshandeh R, Samadi M, Azimifar Z, Schaeffer J, editors. Degrees of separation in social networks. Fourth Annual Symposium on Combinatorial Search; 2011.
- 43.
Pareto V, Page A. Manuale di economia politica (Manual of political economy). Milan, Italy: Societa Editrice Libraia. 1906.
- 44. Dubinsky AJ, Hansen RW. IMPROVING MARKETING PRODUCTIVITY—THE 80/20 PRINCIPLE REVISITED. California Management Review. 1982;25(1):96–105. WOS:A1982PU39400008.
- 45. Cooper L, Kang SY, Bisanzio D, Maxwell K, Rodriguez-Barraquer I, Greenhouse B, et al. Pareto rules for malaria super-spreaders and super-spreading. Nature Communications. 2019;10. WOS:000483305700016. pmid:31477710
- 46.
Wilkinson DM, editor Strong regularities in online peer production. Proceedings of the 9th ACM conference on Electronic commerce; 2008.
- 47.
Zhou Z, Bandari R, Kong J, Qian H, Roychowdhury V, editors. Information resonance on twitter: watching iran. Proceedings of the first workshop on social media analytics; 2010.
- 48. Boehmer J, Tandoc EC. Why we retweet: Factors influencing intentions to share sport news on Twitter. International Journal of Sport Communication. 2015;8(2):212–32.
- 49. Preece J, Maloney-Krichmar D. Online communities: focusing on sociability and usability. Handbook of human-computer interaction. 2003:596–620.
- 50.
Kaufmann M, Kalita J, editors. Syntactic normalization of twitter messages. International conference on natural language processing, Kharagpur, India; 2010.
- 51. Fung IC-H, Zeng J, Chan C-H, Liang H, Yin J, Liu Z, et al. Twitter and Middle East respiratory syndrome, South Korea, 2015: A multi-lingual study. Infection, disease & health. 2018;23(1):10–6. pmid:30479298
- 52.
Hong L, Convertino G, Chi EH, editors. Language matters in twitter: A large scale study. Fifth international AAAI conference on weblogs and social media; 2011.
- 53. Liang H, Shen F. Birds of a schedule flock together: Social networks, peer influence, and digital activity cycles. Computers in Human Behavior. 2018;82:167–76.
- 54. Longley PA, Adnan M, Lansley G. The geotemporal demographics of Twitter usage. Environment and Planning A. 2015;47(2):465–84.
- 55.
Grootendorst M. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:220305794. 2022.
- 56. Yang K-C, Varol O, Hui P-M, Menczer F. Scalable and Generalizable Social Bot Detection through Data Selection. Proceedings of the AAAI Conference on Artificial Intelligence. 2020;34(01):1096–103.
- 57. Vosoughi S, Roy D, Aral S. The spread of true and false news online. Science. 2018;359(6380):1146–+. WOS:000426835900044. pmid:29590045
- 58. Chong M, Park HW. COVID-19 in the Twitterverse, from epidemic to pandemic: information-sharing behavior and Twitter as an information carrier. Scientometrics. 2021;126(8):6479–503. pmid:34188332
- 59. Park H, Biddix JP, Park HW. Discussion, news information, and research sharing on social media at the onset of Covid-19. El Profesional de la Información. 2021;30(4).
- 60. Park HW, Park S, Chong M. Conversations and medical news frames on Twitter: Infodemiological study on COVID-19 in South Korea. Journal of medical internet research. 2020;22(5):e18897. pmid:32325426
- 61.
Uyheng J, Carley KM. Computational Analysis of Bot Activity in the Asia-Pacific: A Comparative Study of Four National Elections. Proceedings of the International AAAI Conference on Web and Social Media. 2021;15(1):727–38.
- 62.
Becchetti L, Castillo C, Donato D, Fazzone A, Rome I, editors. A comparison of sampling techniques for web graph characterization. Proceedings of the Workshop on Link Analysis (LinkKDD’06), Philadelphia, PA; 2006.
- 63.
Datta S, Kargupta H, editors. Uniform data sampling from a peer-to-peer network. 27th International Conference on Distributed Computing Systems (ICDCS’07); 2007: IEEE.
- 64.
Lu J, Li D, editors. Sampling online social networks by random walk. Proceedings of the First ACM International Workshop on Hot Topics on Interdisciplinary Social Networks Research; 2012.