Trends in Social Media : Persistence and Decay

Sitaram Asur Bernardo A. Huberman

Social Computing Lab Social Computing Lab
HP Labs HP Labs
Palo Alto, California, USA Palo Alto, California, USA
[email protected] [email protected]

Gabor Szabo Chunyan Wang

Social Computing Lab Dept. of Applied Physics
HP Labs Stanford University
Palo Alto, California, USA California, USA
[email protected] [email protected]

ABSTRACT At present there is no clear picture of what causes these topics to

Social media generates a prodigious wealth of real-time content become extremely popular, nor how some persist in the public eye
at an incessant rate. From all the content that people create and longer than others. There is considerable evidence that one aspect
share, only a few topics manage to attract enough attention to rise that causes topics to decay over time is their novelty [11]. Another
to the top and become temporal trends which are displayed to users. factor responsible for their decay is the competitive nature of the
The question of what factors cause the formation and persistence of medium. As content starts propagating throught a social network
trends is an important one that has not been answered yet. In this it can usurp the positions of earlier topics of interest, and due to
paper, we conduct an intensive study of trending topics on Twitter the limited attention of users it is soon rendered invisible by newer
and provide a theoretical basis for the formation, persistence and content. Yet another aspect responsible for the popularity of certain
decay of trends. We also demonstrate empirically how factors such topics is the influence of members of the network on the propaga-
as user activity and number of followers do not contribute strongly tion of content. Some users generate content that resonates very
to trend creation and its propagation. In fact, we find that the res- strongly with their followers thus causing the content to propagate
onance of the content with the users of the social network plays a and gain popularity [9].
major role in causing trends.
The source of that content can originate in standard media outlets
or from users who generate topics that eventually become part of
1. INTRODUCTION the trends and capture the attention of large communities. In either
Social media is growing at an explosive rate, with millions of peo- case the fact that a small set of topics become part of the trending
ple all over the world generating and sharing content on a scale set means that they will capture the attention of a large audience
barely imaginable a few years ago. This has resulted in massive for a short time, thus contributing in some measure to the public
participation with countless number of updates, opinions, news, agenda. When topics originate in media outlets, the social medium
comments and product reviews being constantly posted and dis- acts as filter and amplifier of what the standard media produces and
cussed in social web sites such as Facebook, Digg and Twitter, to thus contributes to the agenda setting mechanisms that have been
name a few. thoroughly studied for more than three decades [7] .
This widespread generation and consumption of content has cre- In this paper, we study trending topics on Twitter, an immensely
ated an extremely competitive online environment where different popular microblogging network on which millions of users create
types of content vie with each other for the scarce attention of the and propagate enormous content via a steady stream on a daily ba-
user community. In spite of the seemingly chaotic fashion with sis. The trending topics, which are shown on the main website, rep-
which all these interactions take place, certain topics manage to resent those pieces of content that bubble to the surface on Twitter
attract an inordinate amount of attention, thus bubbling to the top owing to frequent mentions by the community. Thus they can be
in terms of popularity. Through their visibility, this popular top- equated to crowdsourced popularity. We then determine the fac-
ics contribute to the collective awareness of what is trending and at tors that contribute to the creation and evolution of these trends, as
times can also affect the public agenda of the community. they provide insight into the complex interactions that lead to the
popularity and persistence of certain topics on Twitter, while most
others fail to catch on and are lost in the flow.

We first analyze the distribution of the number of tweets across

trending topics. We observe that they are characterized by a strong
log-normal distribution, similar to that found in other networks
such as Digg and which is generated by a stochastic multiplicative
process [11]. We also find that the decay function for the tweets is
mostly linear. Subsequently we study the persistence of the trends
to determine which topics last long at the top. Our analysis reveals immense potential it provides for viral marketing. Due to its huge
that there are few topics that last for long times, while most topics reach, Twitter is increasingly used by news organizations to dis-
break fairly quickly, in the order of 20-40 minutes. Finally, we look seminate news updates, which are then filtered and commented on
at the impact of users on trend persistence times within Twitter. We by the Twitter community. A number of businesses and organiza-
find that traditional notions of user influence such as the frequency tions are using Twitter or similar micro-blogging services to adver-
of posting and the number of followers are not the main drivers tise products and disseminate information to stockholders.
of trends, as previously thought. Rather, long trends are charac-
terized by the resonating nature of the content, which is found to
arise mainly from traditional media sources. We observe that social 4. TWITTER TRENDS DATA
media behaves as a selective amplifier for the content generated by Trending topics are presented as a list by Twitter on their main Twit-
traditional media, with chains of retweets by many users leading to ter.com site, and are selected by an algorithm proprietary to the
the observed trends. service. They mostly consist of two to three word expressions, and
we can assume with a high confidence that they are snippets that
appear more frequently in the most recent stream of tweets than
2. RELATED WORK one would expect from a document term frequency analysis such
There has been some prior work on analyzing connections on Twit-
as TFIDF. The list of trending topics is updated every few minutes
ter. Huberman et al. [5] studied social interactions on Twitter to
as new topics become popular.
reveal that the driving process for usage is a sparse hidden network
underlying the friends and followers, while most of the links rep-
Twitter provides a Search API for extracting tweets containing par-
resent meaningless interactions. Jansen et al. [6] have examined
ticular keywords. To obtain the dataset of trends for this study,
Twitter as a mechanism for word-of-mouth advertising. They con-
we repeatedly used the API in two stages. First, we collected the
sidered particular brands and products and examined the structure
trending topics by doing an API query every 20 minutes. Second,
of the postings and the change in sentiments. Galuba et al. [4]
for each trending topic, we used the Search API to collect all the
proposed a propagation model that predicts which users will tweet
tweets mentioning this topic over the past 20 minutes. For each
about which URL based on the history of past user activity.
tweet, we collected the author, the text of the tweet and the time it
was posted. Using this procedure for data collection, we obtained
Yang and Leskovec [12] examined patterns of temporal behavior
16.32 million tweets on 3361 different topics over a course of 40
for hashtags in Twitter. They presented a stable time series cluster-
days in Sep-Oct 2010.
ing algorithm and demonstrate the common temporal patterns that
tweets containing hashtags follow. There have also been earlier
We picked 20 minutes as the duration of a timestamp after evaluat-
studies focused on social influence and propagation. Agarwal et
ing different time lengths, to optimize the discovery of new trends
al. [1] studied the problem of identifying influential bloggers in the
while still capturing all trends. This is due to the fact that Twit-
blogosphere. They discovered that the most influential bloggers
ter only allows 1500 tweets per search query. We found that with
were not necessarily the most active. Aral et al, [2] have distin-
20 minute intervals, we were able to capture all the tweets for the
guished the effects of homophily from influence as motivators for
trending topics efficiently.
propagation. As to the study of influence within Twitter, Cha et
al. [3] performed a comparison of three different measures of influ-
We noticed that many topics become trends again after they stop
ence - indegree, retweets, and user mentions. They discovered that
trending according to the Twitter trend algorithm. We therefore
while retweets and mentions correlated well with each other, the in-
considered these trends as separate sequences: it is very likely that
degree of users did not correlate well with the other two measures.
the spreading mechanism of trends has a strong time component
Based on this, they hypothesized that the number of followers may
with an initial increase and a trailing decline, and once a topic stops
not a good measure of influence. Recently, Romero and others [9]
trending, it should be considered as new when it reappears among
introduced a novel influence measure that takes into account the
the users that become aware of it later. This procedure split the
passivity of the audience in the social network. They developed an
3468 originally collected trend titles into 6084 individual trend se-
iterative algorithm to compute influence in the style of the HITS al-
quences.
gorithm and empirically demonstrated that the number of followers
is a poor measure of influence.
5. DISTRIBUTION OF TWEETS
3. TWITTER We measured the number of tweets that each topic gets in 20 minute
Twitter is an extremely popular online microblogging service, that intervals, from the time the topic starts trending until it stops, as
has gained a very large user following, consisting of close to 200 described earlier. From this we can sum up the tweet counts over
million users. The Twitter graph is a directed social network, where time to obtain the cumulative number of tweets Nq (ti ) of topic q
each user chooses to follow certain other users. Each user submits for any time frame ti ,
periodic status updates, known as tweets, that consist of short mes- i
sages limited in size to 140 characters. These updates typically X
Nq (ti ) = nq (tτ ), (1)
consist of personal information about the users, news or links to τ =1
content such as images, video and articles. The posts made by a
user are automatically displayed on the user’s profile page, as well where nq (t) is the number of tweets on topic q in time interval
as shown to his followers. A retweet is a post originally made by t. Since it is plausible to assume that initially popular topics will
one user that is forwarded by another user. Retweets are useful for stay popular later on in time as well, we can calculate the ratios
propagating interesting posts and links through the Twitter commu- Cq (ti , tj ) = Nq (ti )/Nq (tj ) for topic q for time frames ti and tj .
nity. Figure 1(a) shows the distribution of Cq (ti , tj )’s over all topics for
four arbitrarily chosen pairs of time frames (nevertheless such that
Twitter has attracted lots of attention from corporations due to the ti > tj , and ti is relatively large, and tj is small).
 3.5 3.5 3.5
3.0 3.0 4
3.0
2.5 2.5 2.5
3
2.0 2.0
Density

Density

Density

Density
2.0
1.5 1.5 2 1.5
1.0 1.0 1.0
1
0.5 0.5 0.5
0.0 0.0 0 0.0

100.2 100.4 100.6 100.8 101 101.2 101.4 101.6 100.4 100.6 100.8 101 101.2 101.4 101.6 100.2 100.4 100.6 100.8 101 100.2 100.4 100.6 100.8 101 101.2 101.4 101.6
Relative tweet count Relative tweet count Relative tweet count Relative tweet count
(a)
−0.2 1.6
1.6
−0.4 1.4
1.4 0.8
−0.6 1.2
Data quantiles

Data quantiles

Data quantiles

Data quantiles
1.2
−0.8 0.6 1.0
1.0
−1.0 0.8
0.8 0.4
−1.2 0.6
0.6
−1.4 0.2 0.4
0.4

−3 −2 −1 0 1 2 3 −2 −1 0 1 2 −3 −2 −1 0 1 2 3 −2 −1 0 1 2
Normal theoretical quantiles Normal theoretical quantiles Normal theoretical quantiles Normal theoretical quantiles
(b)

Figure 1: (a) The densities of the ratios between cumulative tweet counts measured in two respective time frames. From left to right
in the figure, the indices of the time frames between which the ratios were taken are: (2, 10), (2, 14), (4, 10), and (4, 14), respectively.
The horizontal axis has been rescaled logarithmically, and the solid line in the plots shows the density estimates using a kernel
smoother. (b) The Q-Q plots of the cumulative tweet distributions with respect to normal distributions. If the random variables of
the data were a linear transformation of normal variates, the points would line up on the straight lines shown in the plots. The tails
of the empirical distributions are apparently heavier than in the normal case.

These figures immediately suggest that the ratios Cq (ti , tj ) are dis- Here Nq (0) is the initial number of tweets in the earliest time step.
tributed according to log-normal distributions, since the horizontal Taking the logarithm of both sides of Eq. (3),
axes are logarithmically rescaled, and the histograms appear to be
t
Gaussian functions. To check if this assumption holds, consider X
Fig. 1(b), where we show the Q-Q plots of the distributions of ln Nq (t) − ln Nq (0) = ln [1 + γ(s)ξ(s)] (4)
Fig. 1(a) in comparison to normal distributions. We can observe s=1

that the (logarithmically rescaled) empirical distributions exhibit

The RHS of Eq. (4) is the sum of a large number of random vari-
normality to a high degree for later time frames, with the excep-
ables. The central limit theorem states thus that if the random
tion of the high end of the distributions. These 10-15 outliers occur
variables are independent and identically distributed, then the sum
more frequently than could be expected for a normal distribution.
asymptotically approximates a normal distribution. The i.i.d condi-
tion would hold exactly for the ξ(s) term, and it can be shown that
Log-normals arise as a result of multiplicative growth processes
in the presence of the discounting factors (if the rate of decline is
with noise [8]. In our case, if Nq (t) is the number of tweets for
not too fast), the resulting distribution is still normal [11].
a given topic q at time t, then the dynamics that leads to a log-
normally distributed Nq (t) over q can be written as:
In other words, we expect from this model that ln [Nq (t)/Nq (0)]
will be distributed normally over q when fixing t. These quantities
Nq (t) = [1 + γ(t)ξ(t)] Nq (t − 1), (2)
were shown in Fig. 1 above. Essentially, if the difference between
the two times where we take the ratio is big enough, the log-normal
where the random variables ξ(t) are positive, independent and iden-
property is observed.
tically distributed as a function of t with mean 1 and variance σ 2 .
Note that time here is measured in discrete steps (t − 1 expresses
The intuitive explanation for the multiplicative model of Eq. (2)
the previous time step with respect to t), in accordance with our
is that at each time step the number of new tweets on a topic is a
measurement setup. γ(t) is introduced to account for the novelty
multiple of the tweets that we already have. The number of past
decay [11]. We would expect topics to initially increase in popu-
tweets, in turn, is a proxy for the number of users that are aware of
larity but to slow down their activity as they become obsolete or
the topic up to that point. These users discuss the topic on different
known to most users. Since γ(t) is made up of decreasing positive
forums, including Twitter, essentially creating an effective network
numbers, the growth of Nt slows with time.
through which the topic spreads. As more users talk about a par-
ticular topic, many others are likely to learn about it, thus giving
To see that Eq. (2) leads to a log-normal distribution of Nq (t), we
the multiplicative nature of the spreading. The noise term is nec-
first expand the recursion relation:
essary to account for the stochasticity of this process. On the other
t hand, the monotically decreasing γ(t) characterizes the decay in
timeliness and novelty of the topic as it slowly becomes obsolete
Y
Nq (t) = [1 + γ(s)ξ(s)] Nq (0). (3)
s=1 and known to most users, and guarantees that Nq (t) does not grow
 1.2 1.0
100
1.0

0.8

−0.5 0.6 0.8

10

Normalized tweet counts

0.4

0.2

0.6
10−1
10 20 30 40
γ

10−1.5 0.4

10−2 0.2

0.0

100 100.5 101 101.5 0 10 20 30 40

Time (hours) Time (hours)

Figure 3: The number of total tweets on topics in the first

Figure 2: The decay factor γ(t) in time as measured using 48 hours, normalized to 1 so that they can be shown on the
Eq. (5). The log-log plot exhibits that it decreases in a power- same plot. The randomly selected topics were (from left to
law fashion, with an exponent that is measured to be exactly -1 right): “Earnings”, “#pulpopaul”, “Sheen”, “Deuces Remix”,
(the linear regression on the logarithmically transformed data “Isaacs”, “#gmp24”, and “Mac App”.
fits with R2 = 0.98). The fit to determine the exponent was
performed in the range of the solid line next to the function,
which also shows the result of the fit while being shifted lower
for easy comparison. The inset displays the same γ(t) function This suggests that the trending topics featured on Twitter increase
on standard linear scales. their tweet counts linearly in time, and their dynamics is captured
by the multiplicative noise model we discussed above.

unbounded [11]. To check this, we first plotted a few representative examples of the
cumulative number of tweets for a few topics in Fig. 3. It is ap-
To measure the functional form of γ(t), we observe that the ex- parent that all the topics ( selected randomly) show an approximate
pected value of the noise term ξ(t) in Eq. (2) is 1. Thus averaging initial linear growth in the number of tweets.We also checked if this
over the fractions between consecutive tweet counts yields γ(t): is true in general. Figure 4 shows the second discrete derivative of
  the total number of tweets, which we expect to be 0 if the trend
Nq (t)
γ(t) = − 1. (5) lines are linear on average. A positive second derivative would
Nq (t − 1) q mean that the growth is superlinear, while a negative one suggests
The experimental values of γ(t) in time are shown in Fig. 2. It that it is sublinear. We point out that before taking the average of
is interesting to notice that γ(t) follows a power-law decay very all second derivatives over the different topics in time, we divided
precisely with an exponent of −1, which means that γ(t) ∼ 1/t. the derivatives by the average of the total number of tweets of the
given topics. We did this so as to account for the large difference
between the ranges of the number of tweets across topics, since a
6. THE GROWTH OF TWEETS OVER TIME simple averaging without prior normalization would likely bias the
The interesting fact about the decay function γ(t) = 1/t is that results towards topics with large tweet counts and their fluctuations.
it results in a linear increase in the total number of tweets for a The averages are shown in Fig. 4.
topic over time. To see this, we can again consider Eq. (4), and
approximate the discrete sum of random variables with an integral We observe from the figure that when we consider all topics there
of the operand of the sum, and substitute the noise term with its is a very slight sublinear growth regime right after the topic starts
expectation value, hξ(t)i = 1 as defined earlier (this is valid if γ(t) trending, which then becomes mostly linear, as the derivatives data
is changing slowly). These approximations yield the following: is distributed around 0. If we consider only very popular topics
Nq (t)
Z t Z t
1 (that were on the trends site for more than 4 hours), we observe an
ln ≈ ln [1 + γ(τ )] dτ ≈ dτ = ln t. (6) even better linear trend. One reason for this may be that topics that
Nq (0) τ =0 τ =0 τ
trend only for short periods exhibit a concave curvature, since they
In simplifying the logarithm above, we used the Taylor expansion lose popularity quickly, and are removed from among the Twitter
of ln(1 + x) ≈ x, for small x, and also used the fact that γ(τ ) = trends by the system early on.
1/τ as we found experimentally earlier.
These results suggest that once a topic is highlighted as a trend on a
It can be immediately seen then that Nq (t) ≈ Nq (0) t for the range very visible website, its growth becomes linear in time. The reason
of t where γ(t) is inversely proportional to t. In fact, it can be for this may be that as more and more visitors come to the site
easily proven that no functional form for γ(t) would yield a lin- and see the trending topics there is a constant probability that they
ear increase in Nq (t) other than γ(t) ∼ 1/t (assuming that the will also talk and tweet about it. This is in contrast to scenarios
above approximations are valid for the stochastic discrete case). where the primary channel of information flow is more informal.
 0.06 2000

103
Relative curvature of the tweet count

1500

0.04
102.5 1000

Number of topics
500

0.02 102
0

0 2 4 6 8 10

101.5
0.00

101
−0.02
100.5

−0.04
100

100 100.5 101 101.5

−0.06 Number of topic recurrences
0 5 10 15 20
Time (hours)
103.5
2000

Figure 4: The average of the second derivative of the total num- 103 1500

ber of tweets over all topics. For one topic, we first divided the

Frequency
2.5
derivative values by the mean of the tweet counts so as to min- 10 1000

imize the differences between the wide range of topic popular- 500

102

Frequency
ities. The open circles show the derivatives obtained with this 0

0 5 10 15 20

procedure for all topics, while the smaller red dots represent 101.5
Trend duration (hours)

only topics that trended for longer than 4 hours.

101

100.5
In that case we expect that the growth will exhibit first a phase
with accelerated growth and then slow down to a point when no 100
one talks about the topic any more. Content that spreads through
100 100.5 101 101.5
a social network or without external “driving” will follow such a Trend duration (hours)
course, as has been showed elsewhere [10, 12].
Figure 5: (a) The distribution of the number of sequences a
7. PERSISTENCE OF TRENDS trending topic comprises of (b) The distribution of the lengths
An important reason to study trending topics on Twitter is to un- of each sequence. Both graphs are shown in the log-log scale
derstand why some of them remain at the top while others dissi- with the inset giving the actual histograms in the linear scale.
pate quickly. To see the general pattern of behavior on Twitter,
we examined the lifetimes of the topics that trended in our study.
From Fig 5(a) we can see that while most topics occur continu- We first examine the authors who tweet about given trending topics
ously, around 34% of topics appear in more than one sequence. to see if the authors change over time or if it is the same people
This means that they stop trending for a certain period of time be- who keep tweeting to cause trends. When we computed the corre-
fore beginning to trend again. lation in the number of unique authors for a topic with the duration
(number of timestamps) that the topic trends we noticed that cor-
A reason for this behavior may be the time zones that are involved. relation is very strong (0.80). This indicates that as the number of
For instance, if a topic is a piece of news relevant to North Ameri- authors increases so does the lifetime, suggesting that the propaga-
can readers, a trend may first appear in the Eastern time zone, and tion through the network causes the topic to trend.
3 hours later in the Pacific time zone. Likewise, a trend may re-
turn the next morning if it was trending the previous evening, when To measure the impact of authors we compute for each topic the
more users check their accounts again after the night. active-ratio aq as:
N umber of T weets
Given that many topics do not occur continuously, we examined the aq = (7)
N umber of U nique Authors
distribution of the lengths sequences for all topics. In Fig 5(b) we
show the length of the topic sequences. It can be observed that this The correlation of active-ratio with trending duration is as shown in
is a power-law which means that most topic sequences are short Fig 6. We observe that the active-ratio quickly saturates and varies
and a few topics last for a very long time. This could be due to the little with time for any given topic. Since the authors change over
fact that there are many topics competing for attention. Thus, the time with the topic propagation, the correlation between number of
topics that make it to the top (the trend list) last for a short time. tweets and authors is high (0.83).
However, in many cases, the topics return to trend for more time,
which is captured by the number of sequences shown in Fig 5(a), 7.2 Persistence of long trending topics
as mentioned. On Twitter each topic competes with the others to survive on the
trending page. As we now show, for the long trending ones we can
7.1 Relation to authors and activity derive an expression for the distribution of their average length.
 ●

4000
3000 0.1
Trending Duration (mins)

Probability
●
2000

0.01
●
●

●● ● ●●●
●●●
● ●
● ●
1000

●●●
●●● ● ●
●●●●
●●●
●● ● ●
●●●
●●
●●
●●● ●
●● ●
● ●● ●
●
●●
●
●●●
● ● ● ● ●
●
●● ● ● ● ●
●
●
●
●
●●
●
●
●
●●●
● ● ●●● ●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●● ●●● ●● ● ● ●● ●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●●●●●●●
●● ●●● ● ● ● ● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
● ●
● ●●
●●
● ● ● ● ● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●●
●●
●
● ●●●●
●● ●● ●● ●
●
●
●●
●●
●●
●
●
●●
●●●
●●●●●●
●●●●● ●●●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
●●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ● ●●
●
● ●● ●
●●
●
● ●●●
●●
● ● 1E-3
0

10 20 30 40
2 4 6 8 10
Trending Duration
Active Ratio

Figure 6: Relation between the active-ratio and the length of Figure 8: Fit of trending duration to density in log scale. The
the trend across all topics, showing that the active-ratio does straight line suggests an exponential family of the trending time
not vary significantly with time. distribution. The red line gives a fit with an R2 of 0.9112.

F (x) is the cumulative distribution function of the random variable

χ. Given that distribution we can actually determine the threshold
for survival as:
0.12
−1
θ = eF (p)
(10)
From the independence property of the φ, the duration or life time
0.08 of a trending topic, denoted by L, follows a geometric distribution,
Probability

which in the continuum case becomes the exponential distribution.

Thus, the probability that a topic survives in the first k time inter-
0.04
vals and fails in the k + 1 time interval, given that k is large, can be
written as:

0.00

Pr(L = k) = (1 − p)k p (11)

10 15 20 25 30 35 40

Trending Duration

The expected length of trending duration L would thus be:

∞
Figure 7: Distribution of trending times. The black dots repre- X 1 1
sents actual trending data pulled from Twitter, and the red dots < L >= (1 − p)k p · k = −1= −1 (12)
0
p F (log θ)
are the predictions from a geometric distribution with p=0.12.

We considered trending durations for topics that trended for more

We assume that, if the relative growth rate of tweets, denoted by than 10 timestamps on Twitter. The comparison between the ge-
φt = NNt−1
t
, falls below a certain threshold θ, the topic would stop ometric distribution and the trending duration is shown in Fig 7.
In Fig 8 the fit of the trending duration to density in a logarithmic
trending. When we consider long-trending topics, as they grow
scale suggests an exponential function for the trending time. The
in time, they overcome the initial novelty decay, and the γ term
R-square of the fitting is 0.9112.
in equation (3) becomes fairly constant. So we can measure the
change over time using only the random variable ξ as :
8. TREND-SETTERS
We consider two types of people who contribute to trending topics -
the sources who begin trends, and the propagators who are respon-
Nt Nt Nt−1
log φt = log = log − log ' ξt (8) sible for those trends propagating through the network due to the
Nt−1 N0 N0 nature of the content they share.

Since the ξs are independent and identical distributed random vari- 8.1 Sources
ables, φ1 , φ2 , ···φt would be independent with each other. Thus the We examined the users who initiate the most trending topics. First,
probability that a topic stops trending in a time interval s, where s is for each topic we extracted the first 100 users who tweeted about it
large, is equal to the probability that φs is lower than the threshold prior to its trending. The distribution of these authors and the topics
θ, which can be written as: is a power-law, as shown in Fig 9. This shows that there are few
authors who contribute to the creation of many different topics. To
p = Pr(φs < θ) = Pr(log φs < log(θ)) focus on these multi-tasking users, we considered only the authors
(9)
= Pr(ξs < log(θ)) = F (log θ) who contributed to at least five trending topics.
 Author Retweets Topics Retweet-Ratio
105 200000 vovo_panico 11688 65 179.81
150000 cnnbrk 8444 84 100.52
104 100000
keshasuja 5110 51 100.19
LadyGonga 4580 54 84.81
Number of authors

50000

103 0
BreakingNews 8406 100 84.06
0 2 4 6 8 10
MLB 3866 62 62.35
nytimes 2960 59 50.17
102
HerbertFromFG 2693 58 46.43
espn 2371 66 35.92
101 globovision 2668 75 35.57
huffingtonpost 2135 63 33.88
100 skynewsbreak 1664 52 32
100 100.5 101 101.5 el_pais 1623 52 31.21
Number of trending topics initiated stcom 1255 51 24.60
la_patilla 1273 65 19.58
Figure 9: Distribution of the first 100 authors for each trending reuters 957 57 16.78
topic. The log-log plot shows a power-law distribution. The WashingtonPost 929 60 15.48
inset graph gives the actual histogram in the linear scale. bbcworld 832 59 14.10
CBSnews 547 56 9.76
When we consider people who are influential in starting trends on TelegraphNews 464 79 5.87
Twitter, we can hypothesize two attributes - a high frequency of ac- tweetmeme 342 97 3.52
tivity for these users, as well as a large follower network. To eval- nydailynews 173 51 3.39
uate these hypotheses we measured these two attributes for these Table 1: Top 22 Retweeted Users in at least 50 trending topics
authors over these months. each
Frequency: The tweet-rate can effectively measure the frequency identify the authors who are retweeted the most in the trending top-
of participation of a Twitter user. The mean tweet-rate for these ics, we counted the number of retweets for each author on each
users was 26.38 tweets per day, indicating that these authors tweeted topic.
fairly regularly. However, when we computed the correlation of the
tweet-rate with the number of trending topics that they contributed Domination: We found that in some cases, almost all the retweets
to, the result was a weak positive correlation of 0.22. This indicates for a topic are credited to one single user. These are topics that are
that although people who tweet a lot do tend to contribute to the entirely based on the comments by that user. They can thus be said
trending topics, the rate by itself does not strongly determine the to be dominating the topic. The domination-ratio for a topic can be
popularity of the topic. In fact, they happen to tweet on a variety of defined as the fraction of the retweets of that topic that can be at-
topics, many of which do not become trends. We found that a large tributed to the largest contributing user for that topic. However, we
number of them tended to tweet frequently about sporting events observed a negative correlation of −0.19 between the domination-
and players and teams involved. When some sports-related topics ratio of a topic to its trending duration. This means that topics
begin to trend, these users are among the early initiators of them, by revolving around a particular author’s tweets do not typically last
virtue of their high tweet-rate. This suggests that the nature of the long. This is consistent with the earlier observed strong correlation
content plays a strong role in determining if a topic trends, rather between number of authors and the trend duration. Hence, for a
than the users who initate it. topic to trend for a long time, it requires many people to contribute
actively to it.
Audience: When we looked at the number of followers for these
authors, we were surprised to find that they were almost completely Influence: On the other hand, we observed that there were authors
uncorrelated (correlation of 0.01) with the number of trending top- who contributed actively to many topics and were retweeted signif-
ics, although the mean is fairly high (2481) 1 . The absence of cor- icantly in many of them. For each author, we computed the ratio
relation indicates that the number of followers is not an indication of retweets to topics which we call the retweet-ratio. The list of
of influence, similar to observations in earlier work [9]. influential authors who are retweeted in at least 50 trending topics
is shown in Table 1. We find that a large portion of these authors
8.2 Propagators are popular news sources such as CNN, the New York Times and
We have observed previously that topics trend on Twitter mainly ESPN. This illustrates that social media, far from being an alter-
due to the propagation through the network. The main way to prop- nate source of news, functions more as a filter and an amplifier for
agate information on Twitter is by retweeting. 31% of the tweets of interesting news from traditional media.
trending topics are retweets. This reflects a high volume of propa-
gation that garner popularity for these topics. Further, the number 9. CONCLUSIONS
of retweets for a topic correlates very strongly (0.96) with the trend To study the dynamics of trends in social media, we have conducted
duration, indicating that a topic is of interest as long as there are a comprehensive study on trending topics on Twitter. We first de-
people retweeting it. rived a stochastic model to explain the growth of trending topics
and showed that it leads to a lognormal distribution, which is vali-
Each retweet credits the original poster of the tweet. Hence, to dated by our empirical results. We also have found that most topics
1 do not trend for long, and for those that are long-trending, their
This is due to the fact that one of these authors has more than a
million followers persistence obeys a geometric distribution.
When we considered the impact of the users of the network, we
discovered that the number of followers and tweet-rate of users are
not the attributes that cause trends. What proves to be more impor-
tant in determining trends is the retweets by other users, which is
more related to the content that is being shared than the attributes
of the users. Furthermore, we found that the content that trended
was largely news from traditional media sources, which are then
amplified by repeated retweets on Twitter to generate trends.

10. REFERENCES
[1] N. Agarwal, H. Liu, L. Tang, and P. S. Yu. Identifying the
Influential Bloggers in a Community. WSDM’08, 2008.
[2] S. Aral, L. Muchnik, and A. Sundararajan. Distinguishing
influence-based contagion from homophily-driven diffusion
in dynamic networks. Proceedings of the National Academy
of Sciences, 106(51):21544–21549, December 2009.
[3] M. Cha, H. Haddadi, F. Benevenuto, and K. P. Gummadi.
Measuring User Influence in Twitter: The Million Follower
Fallacy. In Fourth International AAAI Conference on
Weblogs and Social Media, May 2010.
[4] W. Galuba, D. Chakraborty, K. Aberer, Z. Despotovic, and
W. Kellerer. Outtweeting the Twitterers - Predicting
Information Cascades in Microblogs. In 3rd Workshop on
Online Social Networks (WOSN 2010), 2010.
[5] B. A. Huberman, D. M. Romero, and F. Wu. Social networks
that matter: Twitter under the microscope. ArXiv e-prints,
December 2008, 0812.1045.
[6] B. J. Jansen, M. Zhang, K. Sobel, and A. Chowdury. Twitter
power: Tweets as electronic word of mouth. J. Am. Soc. Inf.
Sci., 60(11):2169–2188, 2009.
[7] M. E. McCombs and D. L. Shaw. The Evolution of
Agenda-Setting Research: Twenty Five Years in the
Marketplace of Ideas. Journal of Communication, (43
(2)):68–84, 1993.
[8] M. Mitzenmacher. A brief history of generative models for
power law and lognormal distributions. Internet
Mathematics, 1:226–251, 2004.
[9] D. M. Romero, W. Galuba, S. Asur, and B. A. Huberman.
Influence and passivity in social media. In 20th International
World Wide Web Conference (WWW’11), 2011.
[10] G. Szabo and B. A. Huberman. Predicting the popularity of
online content. Commun. ACM, 53(8):80–88, 2010.
[11] F. Wu and B. A. Huberman. Novelty and collective attention.
Proceedings of the National Academy of Sciences of the
United States of America, 104(45):17599–17601, November
2007.
[12] J. Yang and J. Leskovec. Patterns of temporal variation in
online media. In Proceedings of the fourth ACM
international conference on Web search and data mining,
WSDM ’11, pages 177–186, 2011.