Measuring User Inuence in Twitter: The Million Follower Fallacy

Meeyoung Cha

Hamed Haddadi

Fabrcio Benevenuto

Krishna P. Gummadi

Max Planck Institute for Software Systems (MPI-SWS), Germany Royal Veterinary College, University of London, United Kingdom CS Dept., Federal University of Minas Gerais (UFMG), Brazil

Abstract
Directed links in social media could represent anything from intimate friendships to common interests, or even a passion for breaking news or celebrity gossip. Such directed links determine the ow of information and hence indicate a users inuence on othersa concept that is crucial in sociology and viral marketing. In this paper, using a large amount of data collected from Twitter, we present an in-depth comparison of three measures of inuence: indegree, retweets, and mentions. Based on these measures, we investigate the dynamics of user inuence across topics and time. We make several interesting observations. First, popular users who have high indegree are not necessarily inuential in terms of spawning retweets or mentions. Second, most inuential users can hold signicant inuence over a variety of topics. Third, inuence is not gained spontaneously or accidentally, but through concerted effort such as limiting tweets to a single topic. We believe that these ndings provide new insights for viral marketing and suggest that topological measures such as indegree alone reveals very little about the inuence of a user.

Introduction
Inuence has long been studied in the elds of sociology, communication, marketing, and political science (Rogers 1962; Katz and Lazarsfeld 1955). The notion of inuence plays a vital role in how businesses operate and how a society functionsfor instance, see observations on how fashion spreads (Gladwell 2002) and how people vote (Berry and Keller 2003). Studying inuence patterns can help us better understand why certain trends or innovations are adopted faster than others and how we could help advertisers and marketers design more effective campaigns. Studying inuence patterns, however, has been difcult. This is because such a study does not lend itself to readily available quantication, and essential components like human choices and the ways our societies function cannot be reproduced within the connes of the lab. Nevertheless, there have been important theoretical studies on the diffusion of inuence, albeit with radically different results. Traditional communication theory states that
Copyright c 2010, Association for the Advancement of Articial Intelligence (www.aaai.org). All rights reserved.

a minority of users, called inuentials, excel in persuading others (Rogers 1962). This theory predicts that by targeting these inuentials in the network, one may achieve a large-scale chain-reaction of inuence driven by word-ofmouth, with a very small marketing cost (Katz and Lazarsfeld 1955). A more modern view, in contrast, de-emphasizes the role of inuentials. Instead, it posits that the key factors determining inuence are (i) the interpersonal relationship among ordinary users and (ii) the readiness of a society to adopt an innovation (Watts and Dodds 2007; Domingos and Richardson 2001). This modern view of inuence leads to marketing strategies such as collaborative ltering. These theories, however, are still just theories, because there has been a lack of empirical data that could be used to validate either of them. The recent advent of social networking sites and the data within such sites now allow researchers to empirically validate these theories. Moving from theory into practice, we nd that there are many other unanswered questions about how inuence diffuses through a population and whether it varies across topics and time. People have different levels of expertise on various subjects. When it comes to marketing, however, this fact is generally ignored. Marketing services actively search for potential inuencers to promote various items. These inuencers range from cool teenagers, local opinion leaders, all the way to popular public gures. However, the advertised items are often far outside the domain of expertise of these hired individuals. So how effective are these marketing strategies? Can a persons inuence in one area be transferred to other areas? In this paper, we present an empirical analysis of inuence patterns in a popular social medium. Using a large amount of data gathered from Twitter, we compare three different measures of inuence: indegree, retweets, and mentions.1 Focusing on different topics, we examine how the three types of inuential users performed in spreading popular news topics. We also investigate the dynamics of an individuals inuence by topic and over time. Finally, we characterize the precise behaviors that make ordinary individuals gain high inuence over a short period of time.
Indegree is the number of people who follow a user; retweets mean the number of times others forward a users tweet; and mentions mean the number of times others mention a users name.
1

The Twitter dataset used in this paper consists of 2 billion follow links among 54 million users who produced a total of 1.7 billion tweets. We refer readers to our project webpage http://twitter.mpi-sws.org/ for a detailed description of the dataset and our data sharing plan. Our study provides several ndings that have direct implications in the design of social media and viral marketing: 1) Analysis of the three inuence measures provides a better understanding of the different roles users play in social media. Indegree represents popularity of a user; retweets represent the content value of ones tweets; and mentions represent the name value of a user. Hence, the top users based on the three measures have little overlap. 2) Our nding on how inuence varies across topics could serve as a useful test for answering how effective advertisement in Twitter would be if one is to employ inuential users. Our analysis shows that most inuential users hold signicant inuence over a variety of topics. 3) Ordinary users can gain inuence by focusing on a single topic and posting creative and insightful tweets that are perceived as valuable by others, as opposed to simply conversing with others. These ndings provide new insights for viral marketing. The rst nding in particular indicates that indegree alone reveals little about the inuence of a user. This has been coined the million follower fallacy by Avnit (Avnit 2009), who pointed to anecdotal evidence that some users follow others simply for etiquetteits polite to follow someone whos following youand many do not read all the broadcast tweets. We have empirically demonstrated that having a million followers does not always mean much in the Twitter world. Instead, we claim that it is more inuential to have an active audience who retweets or mentions the user.

Dening inuence on Twitter

We start by reviewing studies of diffusion of inuence and related work on inuence propagation on Twitter.

Background
There are a number of conicting ideas and theories about how trends and innovations get adopted and spread. The traditional view assumes that a minority of members in a society possess qualities that make them exceptionally persuasive in spreading ideas to others. These exceptional individuals drive trends on behalf of the majority of ordinary people. They are loosely described as being informed, respected, and well-connected; they are called the opinion leaders in the two-step ow theory (Katz and Lazarsfeld 1955), innovators in the diffusion of innovations theory (Rogers 1962), and hubs, connectors, or mavens in other work (Gladwell 2002). The theory of inuentials is intuitive and compelling. By identifying and convincing a small number of inuential individuals, a viral campaign can reach a wide audience at a small cost. The theory spread well beyond academia and has been adopted in many marketing businesses, e.g., RoperASW and Tremor (Gladwell 2002; Berry and Keller 2003).

In contrast, a more modern view of information ow emphasizes the importance of prevailing culture more than the role of inuentials. Some researchers have reasoned that people in the new information age make choices based on the opinions of their peers and friends, rather than by inuentials (Domingos and Richardson 2001). These researchers argued that direct marketing through inuentials would not be as protable as using network-based advertising such as collaborative ltering. The traditional inuentials theory has also been criticized because its information ow process does not take into account the role of ordinary users. In order to compare the role of inuentials and ordinary users, researchers have developed a series of simulations, in which information ows freely between users, and a user adopts an innovation when he is inuenced by more than a threshold of the sample population (Watts and Dodds 2007). Inuentials were dened as those in the top 10% of inuence distribution. The simulation showed that inuentials initiated more frequent and larger cascades than average users, but they were neither necessary nor sufcient for all diffusions, as suggested in the traditional theory. Moreover, in homogeneous networks, inuentials were no more successful in running long cascades than ordinary users. This means that a trends success depends not on the person who starts it, but on how susceptible the society is overall to the trend (Watts 2007). In fact, a trend can be initiated by any one, and if the environment is right, it will spread. Therefore, Watts dubbed early adopters or opinion leaders accidental inuentials. The above competing ideas have remained as hypotheses for several reasons. First is the lack of data that could be used to empirically test them. Although there exist a handful of empirical studies on word-of-mouth inuence (Leskovec, Adamic, and Huberman 2007; Cha, Mislove, and Gummadi 2009), no work has been conducted on the relative order of inuence among individuals. A second issue is the variety of ways that inuence has been dened (Watts 2007; Goyal, Bonchi, and Lakshmanan 2010). It has been unclear what exactly inuence means. Finally, decades have passed since the inuentials theory appeared. Even if the theory was reasonably accurate when it was proposed, things have changed and now we have much more variability in the ow of inuence. In particular, online communities have become a signicant way we receive new information and inuence in such communities needs to be explored. In this paper, we investigate the notion of inuence using a large amount of data collected from a popular social medium, Twitter.

Measuring inuence on Twitter

The Merriam-Webster dictionary denes inuence as the power or capacity of causing an effect in indirect or intangible ways. Despite the large number of theories of inuence in sociology, there is no tangible way to measure such a force nor is there a concrete denition of what inuence means, for instance, in the spread of news. In this paper, we analyze the Twitter network as a news spreading medium and study the types and degrees of inuence within the network. Focusing on an individuals potential to lead others to engage in a certain act, we highlight

three interpersonal activities on Twitter. First, users interact by following updates of people who post interesting tweets. Second, users can pass along interesting pieces of information to their followers. This act is popularly known as retweeting and can typically be identied by the use of RT @username or via @username in tweets. Finally, users can respond to (or comment on) other peoples tweets, which we call mentioning. Mentioning is identied by searching for @username in the tweet content, after excluding retweets. A tweet that starts with @username is not broadcast to all followers, but only to the replied user. A tweet containing @username in the middle of its text gets broadcast to all followers. These three activities represent the different types of inuence of a person: 1. Indegree inuence, the number of followers of a user, directly indicates the size of the audience for that user. 2. Retweet inuence, which we measure through the number of retweets containing ones name, indicates the ability of that user to generate content with pass-along value. 3. Mention inuence, which we measure through the number of mentions containing ones name, indicates the ability of that user to engage others in a conversation.

Related work on Twitter

Several recent efforts have been made to track inuence on Twitter. The Web Ecology Project tracked 12 popular Twitter users for a 10-day period and grouped a users inuence into two types: conversation-based and contentbased (Leavitt et al. 2009). This work concluded that news media are better at spreading content, while celebrities are better at simply making conversation. Our work extends their notion of inuence and uses extensive data to further examine the spatial and temporal dynamics of inuence. More recently, a PageRank-like measure has been proposed to quantify inuence on Twitter (Weng et al. 2010). The authors found high link reciprocity (72%) from a nonrandom sample of 6,748 Singapore-based users, and argued that high reciprocity is indicative of homophily. They then exploited this fact in computing a user inuence rank. Our study, however, contradicts the observation about high reciprocity; near-complete data of Twitter shows low reciprocity (10%). Thus, we predict that social links on Twitter represent an inuence relationship, rather than homophily. Accordingly, we ask what are the different activities on Twitter that represent inuence of a user and to what extent a persons inuence varies across tweet topic and time.

us to gather large amounts of data. We used the Twitter API to gather information about a users social links and tweets. We launched our crawler in August 2009 for all user IDs ranging from 0 to 80 million. We did not look beyond 80 million, because no single user in the collected data had a link to a user whose ID was greater than that value. Out of 80 million possible IDs, we found 54,981,152 in-use accounts, which were connected to each other by 1,963,263,821 social links. We gathered information about a users follow links and all tweets ever posted by each user since the early days of the service. In total, there were 1,755,925,520 tweets. Nearly 8% of all user accounts were set private, so that only their friends could view their tweets. We ignore these users in our analysis. The social link information is based on the nal snapshot of the network topology at the time of crawling and we do not know when the links were formed. The network of Twitter users comprises a single disproportionately large connected component (containing 94.8% of users), singletons (5%), and smaller components (0.2%). The largest component contains 99% of all links and tweets. Our goal is to explore inuence of users, hence we focus on the largest component of the network, which is conceptually a single interaction domain for users. Because it is hard to determine inuence of users who have few tweets, we borrowed the concept of active users from the traditional media research (Levy and Windhal 1985) and focused on those users with some minimum level of activity. We ignored users who had posted fewer than 10 tweets during their entire lifetime. We also ignored users for whom we did not have a valid screen name, because this information is crucial in identifying the number of times a user was mentioned or retweeted by others. After ltering, there were 6,189,636 users, whom we focus on in the remainder of this paper. To measure the inuence of these 6 million users, however, we looked into how the entire set of 52 million users interacted with these active users.

Methodology for comparing user inuence

For each of the 6 million users, we computed the value of each inuence measure and compared them. Rather than comparing the values directly, we used the relative order of users ranks as a measure of difference. In order to do this, we sorted users by each measure, so that the rank of 1 indicates the most inuential user and increasing rank indicates a less inuential user. Users with the same inuence value receive the average of the rank amongst them (Buck 1980). Once every user is assigned a rank for each inuence measure, we are ready to quantify how a users rank varies across different measures and examine what kinds of users are ranked high for a given measure. We used Spearmans rank correlation coefcient (xi yi )2 (1) N3 N as a measure of the strength of the association between two rank sets, where xi and yi are the ranks of users based on two different inuence measures in a dataset of N users. Spearmans rank correlation coefcient calculation is a nonparametric test; the coefcient assesses how well an arbi=1 6

Characteristics of the top inuentials

We describe how we collected the Twitter data and present the characteristics of the top users based on three inuence measures: indegree, retweets, and mentions.

Dataset
We asked Twitter administrators to allow us to gather data from their site at scale. They graciously white-listed the IP address range containing 58 of our servers, which allowed

trary monotonic function could describe the relationship between two variables, without making any other assumptions about the particular nature of the relationship between the variables. Our inclusive and complete dataset guarantees reliability of the correlation estimates. The closer is to +1 or 1, the stronger the likely correlation. A perfect positive correlation is +1 and a perfect negative correlation is 1.

Indegree 27.4 3 7.1 Retweets 26.4 6.7 23.8 Mentions 5.6

Comparing three measures of user inuence

To see what kinds of users are the most inuential, we visited the Twitter pages of the top-20 users based on each measure. The top inuentials The most followed users span a wide variety of public gures and news sources. They were news sources (CNN, New York Times), politicians (Barack Obama), athletes (Shaquille ONeal), as well as celebrities like actors, writers, musicians, and models (Ashton Kutcher, Britney Spears). As the list suggests, indegree measure is useful when we want to identify users who get lots of attention from their audience through one-on-one interactions, i.e., the audience is directly connected to inuentials. The most retweeted users were content aggregation services (Mashable, TwitterTips, TweetMeme), businessmen (Guy Kawasaki), and news sites (The New York Times, The Onion). They are trackers of trending topic and knowledgeable people in different elds, whom other users decide to retweet. Unlike indegree, retweets represent inuence of a user beyond ones one-to-one interaction domain; popular tweets could propagate multiple hops away from the source before they are retweeted throughout the network. Furthermore, because of the tight connection between users as suggested in the triadic closure (Granovetter 1973), retweeting in a social network can serve as a powerful tool to reinforce a messagefor instance, the probability of adopting an innovation increases when not one but a group of users repeat the same message (Watts and Dodds 2007). The most mentioned users were mostly celebrities. Ordinary users showed a great passion for celebrities, regularly posting messages to them or mentioning them, without necessarily retweeting their posts. This indicates that celebrities are often in the center of public attention and celebrity gossip is a popular activity among Twitter users. If retweets represent a citation of another users content, mentions represent a public response to another users tweetthe focus of a tweet is on content for retweets, while the focus is on the replied user for mentions. This can be conrmed from the usage of conventions in tweets: 92% of tweets that had a RT or via marker contained a URL and 97% of them also contained the @username eld. This means that retweets are about the content (indicated by the embedded URL) and that people typically cite the authentic source when they retweet. However, fewer than 30% of tweets that were classied as mentions contained any URL, indicating that a mention is more identity-driven. Across all three measures, the top inuentials were generally recognizable public gures and websites. Interestingly, we saw marginal overlap in these three top lists. These top20 lists only had 2 users in common: Ashton Kutcher and Puff Daddy. The top-100 lists also showed marginal over-

Figure 1: Venn diagram of the top-100 inuentials across measures: The chart is normalized so that the total is 100%. lap, as shown in Figure 1, indicating that the three measures capture different types of inuence. Relative inuence ranks In order to investigate how the three measures correlate, we compared the relative inuence ranks of all 6 million users (Table 1). We see a moderately high correlation (above 0.5) across all pairs. However, the high correlation appears to be an artifact of the tied ranks among the least inuential users, e.g., many of the least connected users also received zero retweet and mention. To avoid this bias, we focused on the set of relatively popular users. We considered users in the top 10th and 1st percentiles based on indegree, in the hope that users who get retweeted or mentioned must have some followers. Table 1: Spearmans rank correlation coefcients Correlation All Top 10% Top 1% Indegree vs retweets 0.549 0.122 0.109 Indegree vs mentions 0.638 0.286 0.309 Retweets vs mentions 0.580 0.638 0.605 After this ltering step, the top users showed a strong correlation in their retweet inuence and mention inuence. Sampling the top users based on retweets or mentions leads to similar results. This means that, in general, users who get mentioned often also get retweeted often, and vice versa. Indegree, however, was not related to the other measures. We conclude that the most connected users are not necessarily the most inuential when it comes to engaging ones audience in conversations and having ones messages spread. Discussion of methodology Normalizing retweets and mentions by total tweets would yield a different measure of inuence, which might have led to very different results. When we tried normalizing the data, we identied local opinion leaders as the most inuential. However, normalization failed to rank users with the highest sheer number of retweets as inuential. Therefore, in this paper, we use the sheer number of retweets and mentions without normalizing these values by the total tweets of a user. Other measures such as the number of tweets and outdegree (i.e., the number of people a user follows) were not found to be useful, because they identied robots and spammers as the most inuential, respectively. Therefore, we do not use these measures.

Topic Iran Swine Jackson

Table 2: Summary information of the three major topics events studied Period Keywords Users Tweets Jun 11Aug 10 #iranelection, names of politicians 302,130 1,482,038 May 3July 2 Mexico u, H1N1, swine 239,329 495,825 Jun 25Aug 24 Michael Jackson, #mj 610,213 1,418,356

Audience 22,177,836 20,977,793 23,550,211

Finally, we calculated indegree based on the snapshot of the network at the time of crawling in 2009, because we do not know the time when each follow link was formed. For the calculation of retweets and mentions, however, we used longitudinal data (i.e., since the beginning of the Twitter service in 2006). This difference could have resulted in a weaker correlation between indegree and the other inuence measures (Table 1). Nevertheless, nearly three quarters of Twitter users joined in 2009, suggesting that the effect on the correlation would have been minimal.

from diverse genres. These properties make this group ideal for studying how a users inuence varies across vastly different topic genres.

Distribution of the inuence ranks

To get a measure of inuence for a given topic, we count only the retweets and mentions a user spawned on the given topic. Because indegree is invariant across topics, we do not use the measure in this analysis. Before investigating the dynamics of inuence across topics, we rst need to understand the high-level characteristics of user inuence, such as the degree to which users inuence can differ. Figure 2 displays the inuence ranks of users based on retweets and mentions across the Iran, Swine, and Jackson topics. The inuence ranks are calculated for each topic and a users rank may differ for different topics. The plots show a straight line on a log-log plot, a property that is referred to as the power-law characteristic. The power-law pattern is indicative of the fact that users degree of inuence can differ by orders of magnitude: the top inuentials were retweeted or mentioned disproportionately more times than the majority of users. This suggests that utilizing top inuentials has a great potential payoff in marketing strategy.
10000 Number of retweets 1000 100 10 1 0 1 10 100 1000 User ranks 10000 Iran Swine Jackson

Does inuence hold across different topics?

Having dened three different measures of inuence on Twitter, we now investigate whether a users inuence varies by topic genres. In order to investigate this, we rst need to nd users who tweeted about a diverse set of topics.

Methodology for identifying target topics

To nd as many users to monitor as possible, we picked three of the most popular topics in 2009 that were considered engaging and revolutionary in Twitter2 : the Iranian presidential election, the outbreak of the H1N1 inuenza, and the death of Michael Jackson. While all three topics were popular, they span political, health, and social genres. To extract tweets relevant to these events, we rst identied the set of keywords describing the topics by consulting news websites and informed individuals. Using the selected keywords, we then identied relevant tweets by searching for the keywords in the tweet dataset. Table 2 displays the keywords and the total number of users and tweets for each topic. We focused on a period of 60 days starting from one day prior to the start of each event. We limited the study duration because popular keywords were typically hijacked by spammers after certain time. The table also shows the total number of users who received any tweet on the topic (termed audience). Each topic reached an audience of over 20 million, indicating that over 40% of users in Twitter were aware of at least one of the three topics. Among the users who tweeted about any of these topics, fewer than 2% discussed all three topics. Although 2% is a small fraction, this set contains 13,219 users, which is a large enough sample size for statistical analysis. These users were generally well connected; they had on average 2,037 followers, and together reached an audience of 16 million. Furthermore, none of these users were dedicated to these three topics; no user had dedicated more than 60% of their tweets to these topics. This means that the set of 13,219 users is enthusiastic about sharing thoughts on popular news topics
The top Twitter trends identied by the Twitter Research team are listed at http://tinyurl.com/yb4965e.
2

(a) Retweet inuence ranks

Number of mentions 1000 100 10 1 0 1 10 100 1000 User ranks 10000 Iran Swine Jackson

(b) Mention inuence ranks

Figure 2: Distribution of user ranks for a given topic

Variation of a users inuence across topics

Ranks (Swine, Jackson)

100 80 60 40 20 0 0

Next, in order to examine how volatile a users rank is across different topics, we compare the relative order of inuence ranks across topics in Table 3. For the same reason we previously discussed in Table 1, we ignore the least popular users who have tied ranks and focus on the set of relatively popular users, as measured by indegree. The top 10th and 1st percentiles included 1,322 and 132 users, respectively. Table 3: Spearmans rank correlation coefcients over topics Retweet Mentions Topics Top 10% 1% Top 10% 1% Iran vs. Swine 0.54 0.62 0.59 0.68 Iran vs. Jackson 0.48 0.54 0.59 0.63 Swine vs. Jackson 0.55 0.50 0.80 0.68 The rank correlation is generally high (above 0.5) and gets stronger for the top 1% of users. Mentions show an even stronger correlation across topics than retweets. This means that a popular user who is good at spawning mentions from others can do so over a wide range of topics, more easily than when she is retweeted over diverse topics. Among topic pairs, Swine and Jackson showed the highest correlation for the top 10% of well-connected users for both the retweet and mention inuence. This is perhaps due to the more social nature of these two events, which differs from the Iran topic where special interest groups like politicians and bloggers played a major role. Given that the rank correlation gets stronger for users with high indegree, we investigated how widely the ranks of the top listed users change by topic. Figure 3 shows the retweet ranks of the top users in the Iran topic against the relative retweet ranks of the same users in the Swine and Jackson topics. It is clear from the gure that the top 5 users in the Iran topic retained their relative ranks in the other two topics. These users were Mashable, CnnBrk, TweetMeme, Time, and BreakingNews, who were all in the category of authoritative news sources and content trackers. The mention inuence also showed a similar trend (not shown here); most inuentials ranked consistently high amongst different topics. And we have observed this trend not only for the hyper-inuentials: moderate inuentials like opinion leaders and evangelists also had consistent inuence ranks over diverse topics, as shown in Table 3. Our ndings about the highly skewed ability of users to inuence others (Figure 2) and the strong correlation in a users inuence rank across different topics (Table 3) together lead to two interesting conclusions. First, most inuential users hold signicant inuence over a variety of topics. This means that local opinion leaders and highly popular gures could indeed be used to spread information outside their area of expertise. In fact, new advertisement campaigns have recently been launched that insert advertisement links into a popular persons tweet (Fiorillo 2009). Second, the power-law trend in the difference among inuence of individuals indicates that it is substantially more effective to target the top inuentials than to employ a massive number of non-popular users in order to kick start a viral campaign.

Swine Jackson

10

15

20

Ranks (Iran)

Figure 3: Users retweet inuence ranks over topics

The rise and fall of inuentials over time

Many factorssocial, political, and economicalaffect popularity and inuence of individuals and organizations. In online social media, such dynamics is facilitated by easy entry of competition. It only costs 140 characters to generate a tweet for any user. Likewise, it will be challenging for inuentials to maintain their status when many emerging local opinion leaders and evangelists enter the arena. Here we analyze the dynamics of individuals inuence over time in two ways. First, we track the popularity of top inuentials over a long term period and check how well they maintain their ranks. Second, we focus on users who increased their inuence in a specic topic over a short time period, in order to understand what behaviors make ordinary individuals inuential.

Maintaining engagement of the top inuentials

Out of all 6 million active users in the Twitter network, we picked the top 100 users based on each of the three measures: indegree, retweets, and mentions. We used all the tweets ever posted on Twitter in identifying these inuentials. Due to the overlap we discussed in Figure 1, there were 233 distinct users, whom we call all-time inuentials. In order to see how the inuence of these all-time inuentials varied over time, we tracked their inuence scores over an 8 month period from January to August 2009. To get a time-varying measure of inuence, we counted the number of retweets and mentions the all-time inuentials spawned in every 15 days over the 8 month period. Because we only know the indegree information based on the nal snapshot of the Twitter network, we do not use this measure. For each user, we computed a single explanatory variable P: the probability that a random tweet posted on Twitter during a 15 day period is a retweet (or a mention) of that user. Normalizing by the total number of tweets posted on Twitter is essential to cancel out any variable effect on the data and allows the underlying characteristics of the data sets to be compared. For instance, because the Twitter network quadrupled over time in terms of the registered users, the total volume of tweets merely increased over time. Hence, if we didnt normalize the results, the trend wouldnt be interesting. Google similarly normalizes the data when analyzing their search trends (Ginsberg et al. 2009). Figure 4 displays the time evolution of the normalized retweets and mentions of the 233 all-time inuential users.

0.0015 Retweet probability, P 0.0012 0.0009 0.0006 0.0003 0 Jan Feb Mar Apr May Jun Time (year 2009) Jul Aug Top 10 users Top 11100 Top 101233 Mention probability, P

0.004 0.003 0.002 0.001 0 Jan Feb Mar Apr May Jun Time (year 2009) Jul Aug Top 10 users Top 11100 Top 101233

(a) Retweet inuence

(b) Mention inuence

Figure 4: The temporal evolution of retweets and mentions for the all-time inuential users. For each data point, the error bars are centered on the average retweet (or mention) probability, and they extend up and down by two standard deviations. Although the values of P appear small, they account for a large volume of retweets and mentions: a single most popular inuential user spawned up to 20,000 retweets and 50,000 mentions over a random 15 day period. In order to capture the trend in more detail, we classied the inuentials into three groups based on indegree: the top 10 users, who were mostly the mainstream news sources; the next 90 users, who were mainly celebrities; and the rest, who were a mixed group of public gures and opinion leaders that competed with the traditional mass media. The evolution of the retweet probability in Figure 4(a) shows that all three groups mildly increased their inuence over time. Large error ranges for the top 10 users indicate a much wider variability in the popularity of this group. The growth, however, is marginal for all three groups. We conjecture that the marginal increase is due to the limited number of tweets users post a day. Broadcasting too many tweets puts even popular users at a risk of being classied as spammers. Hence, Twitter users should moderate the number of broadcast tweets in order to avoid crossing their followers information processing limit. The evolution of the mention probability in Figure 4(b) shows distinct patterns for the three groups. The top 10 users fell in popularity over time; the next group had consistent inuence scores over time; and the last group (the least connected among the three) increased their popularity over time. Users in the last group, surprisingly, spawned on average more mentions than the top 100 users. While this trend is counter-intuitive at rst, the differences between mentions and retweets can explain the trend. The mainstream news organizations in the rst group are retweeted the most, but they are not mentioned the most. This is because their names come up mostly when their content get retweeted; it is hard for media sources to engage users with their identities alone. The second group, comprised of celebrities, is more often mentioned than retweeted because of their name value. Their tweets also get retweeted, when inuence is transferred across topics (Table 3). Evangelists in the last group successfully increased their inuence. While many factors could explain this phenomenon, our manual inspection revealed that these users put signicant efforts in conversing with others (e.g., replying to their audience). In a sense, they need self-advertisement the most, because mass media and celebrities have many other on- and off-line channels of to promote themselves. While our ndings provide an interesting view of how different groups of people maintain their popularity, we should also emphasize that our analysis is in retrospect but not causal. These ndings are based on the set of users who ultimately became popular. We also mention that all inuentials put efforts in posting creative and interesting tweets, as shown from the high correlation in their retweet ranks.

Rising inuence of the ordinary users

Finally we examined the users who increased their inuence over a short time period to understand what behaviors make ordinary users inuential. We focused on the set of users who talked about only one news topic, out of the three news topics in Table 2, and picked the top 20 users for each news topic based on indegree. We call these users topical inuentials. The topical inuentials included dedicated accounts like iranbaan, oxfordgirl, and TM Outbreak who suddenly became popular over the course of the event. These users were literally unheard-of at the beginning of the news events and didnt receive any retweets or mentions prior to the relevant news event. The list also included users like kevinrose (the founder of digg.com) and 106andpark (BET.coms music video site) who were already popular on a specic topic, and used the news events to extend their popularity. We tracked the inuence scores of the 60 topical inuentials over the 8 month in 2009. Again, we computed the variable P as the probability that a random tweet posted in a 15 day period is a retweet (or a mention) to that user. Figure 5 displays the temporal evolution of inuence for the topical inuentials. Overall, the inuence scores are much lower than that of the global inuentials in Figure 4. This is expected since topical inuentials had 3 to 180 times fewer followers than the global inuentials. Inuentials on Swine u experienced relatively stable inuence scores for both retweets and mentions. This is because no single day involved a catastrophic event due to Swine u. Inuentials on the Iran election increased their

0.0005 Retweet probability, P 0.0004 0.0003 0.0002 0.0001 0 Jan Feb Mar Apr May Jun Time (year 2009) Jul Aug Iran Jackson Swine Mention probability, P 0.0006 Iran Jackson Swine

0.0004

0.0002

0 Jan Feb Mar Apr May Jun Time (year 2009) Jul Aug

(a) Retweet inuence

(b) Mention inuence

Figure 5: The temporal evolution of retweets and mentions for the topical inuential users retweet inuence substantially during the peak of the election in June and July, but they did not spawn many mentions. Our data shows that these inuentials actively spread information about protests and controversial news. Inuentials on Michael Jackson experienced a mild increase in their retweet inuence in June, whereas their mention inuence surged during the same time period. Our manual inspection revealed that users who limit their tweets to a single topic showed the largest increase in their inuence scores.

References
Avnit, A. 2009. The Million Followers Fallacy, Internet Draft, Pravda Media. http://tinyurl.com/nshcjg. Berry, J., and Keller, E. 2003. The Inuentials: One American in Ten Tells the Other Nine How to Vote, Where to Eat, and What to Buy. Free Press. Buck, W. 1980. Tests of Signicance for Point-Biserial Rank Correlation Coefcients in the Presence of Ties. Biometrical Journal 22(2):153158. Cha, M.; Mislove, A.; and Gummadi, K. P. 2009. A MasurementDriven Analysis of Information Propagation in the Flickr Social Network. In WWW. Domingos, P., and Richardson, M. 2001. Mining the Network Value of Customers. In ACM SIGKDD. Fiorillo, J. 2009. Twitter Advertising: Pay-Per-Tweet, Wikinomics. http://tinyurl.com/d89uu9. Ginsberg, J.; Mohebbi, M. H.; Patel, R. S.; Brammer, L.; Smolinski, M. S.; and Brilliant, L. 2009. Detecting Inuenza Epidemics Using Search Engine Query Data. Nature 457. Gladwell, M. 2002. The Tipping Point: How Little Things Can Make a Big Difference. Back Bay Books. Goyal, A.; Bonchi, F.; and Lakshmanan, L. V. S. 2010. Learning Inuence Probabilities In Social Networks. In ACM WSDM. Granovetter, M. 1973. The Strength of Weak Ties. American Journal of Sociology 78(6). Katz, E., and Lazarsfeld, P. 1955. Personal Inuence: The Part Played by People in the Flow of Mass Communications. New York: The Free Press. Leavitt, A.; Burchard, E.; Fisher, D.; and Gilbert, S. 2009. The Inuentials: New Approaches for Analyzing Inuence on Twitter. Web Ecology Project, http://tinyurl.com/lzjlzq. Leskovec, J.; Adamic, L. A.; and Huberman, B. A. 2007. The Dynamics of Viral Marketing. ACM Trans. on the Web (TWEB). Levy, M., and Windhal, S. 1985. Media Gratications Research, Chapter The Concept of Audience Activity. Sage. Rogers, E. M. 1962. Diffusion of Innovations. Free Press. Watts, D., and Dodds, P. 2007. Inuentials, Networks, and Public Opinion Formation. Journal of Consumer Research. Watts, D. 2007. Challenging the Inuentials Hypothesis. Word of Mouth Marketing Association. Weng, J.; Lim, E.-P.; Jiang, J.; and He, Q. 2010. TwitterRank: Finding Topic-Sensitive Inuential Twitterers. In ACM WSDM.

Conclusions
This paper analyzed the inuence of Twitter users by employing three measures that capture different perspectives: indegree, retweets, and mentions. We found that indegree represents a users popularity, but is not related to other important notions of inuence such as engaging audience, i.e., retweets and mentions. Retweets are driven by the content value of a tweet, while mentions are driven by the name value of the user. Such subtle differences lead to dissimilar groups of the top Twitter users; users who have high indegree do not necessarily spawn many retweets or mentions. This nding suggests that indegree alone reveals very little about the inuence of a user. Focusing on retweets and mentions, we studied the dynamics of inuence across topics and time. Our spatial analysis showed that most inuential users can hold signicant inuence over a variety of topics. The top Twitter users had a disproportionate amount of inuence, which was indicated by a power-law distribution. Our temporal analysis identied how different types of inuentials interact with their audience. Mainstream news organizations consistently spawned a high level of retweets over diverse topics. In contrast, celebrities were better at inducing mentions from their audience. This is because the name value of the mention inuentials helped them get responses from others, rather than any inherent value in the content they posted. Finally, we found that inuence is not gained spontaneously or accidentally, but through concerted effort. In order to gain and maintain inuence, users need to keep great personal involvement. This could mean that inuential users are more predictable than suggested by theory (Watts 2007), shedding light on how to identify emerging inuential users.