Sie sind auf Seite 1von 7

Socio-Temporal influence scoring in Twitter∗


[Extended Abstract]
Khalid Azam Mohar Guha
khalidazam@yahoo.com guhamohar@gmail.com

ABSTRACT With the mainstream adoption of social networks among


In the world of social media marketing, influence is the pre- the masses, there has been an upsurge of interest among
vailing currency. The goal therefore is to find influential researchers and firms in determining how to quantify influ-
users within a network who can be tapped to become the ence. This interest is further leveraged by the realization
brand ambassadors of a product or agents of a social change. within the marketing community that given a century of
In this paper, we propose a modified PageRank model to as- advances in marketing and persuasion,”word of mouth” still
sign influence scores to nodes of a graph in Twitter. Here, remains the most effective form of promotion [1]. Once in-
we use an “infection spread” model which is a continuum ver- fluence is converted into a metric, it has the potential to
sion of the Information Cascade model widely used to study automate”word of mouth” marketing. There are many other
information diffusion within a social network. For each user uses of influence as a metric besides marketing. Social net-
in the network, we calculate the overall susceptibility of all works like Facebook and Twitter have played an instrumen-
his followers to be influenced or infected with a message that tal role in bringing about tectonic social changes such as the
the user tweets. We integrate this infection spread parame- Egyptian and Tunisian revolutions [2][3] and in placing an
ter into the modified PageRank that also takes into account international spotlight on a relatively unknown warlord such
the social graph characteristics of the user and calculate the as Joseph Kony [4]. Unlike the Arab Spring revolutions, in
influence score of the user. which activists used social networks as a means to an end,
the Kony2012 campaign [5] was designed from the ground
Categories and Subject Descriptors up with an influence maximization strategy in which select
highly influential nodes (celebrities, politicians, etc) were
H.4 [Information Systems Applications]: Miscellaneous;
prevailed upon to spread the message. This resulted in a
D.2.8 [Software Engineering]: Metrics—complexity mea-
cascade of message propagation culminating in a 100 mil-
sures, performance measures
lion views of the campaign’s video within a matter of days
[6]. In this paper, we explore the formulation of influence
General Terms as a metric by quantifying it as a social influence score that
Theory can be assigned to each node in a social network. We use
sample data from Twitter collected between 2008 and 2009
Keywords [7][8][9]. Based on existing research as well as our intuition,
Influence score, Pagerank, Twitter, Infection model, Influ- we posit that the influence wielded by a user in the social
ence maximization network is a function of the following attributes:

1. INTRODUCTION • Influence based on user’s social graph


1.1 Motivation (a) The number of followers of the user.
∗(Does NOT produce the permission block, copyright (b) The influence and quality of the followers of the
user.
information nor page numbering). For use with
ACM PROC ARTICLE-SP.CLS. Supported by ACM. (c) The strategic position of the user within the social
†A full version of this paper is available as Author’s Guide to graph.
Preparing ACM SIG Proceedings Using LATEX2 and BibTeX
at www.acm.org/eaddress.htm
• Influence based on user’s activity
(a) The activity level in terms of the number or fre-
quency of tweets sent by the user.
(b) How many of the user’s tweets have been retweeted
by their followers and other users of the network.

• Influence based on user’s expertise or credentials


(a) The correlation of the user’s tweet to his or her
subject matter expertise or topic attribute.
We then combine these influence attributes using a modified works. They propose a Topical Affnity Propagation (TAP)
Page Rank and Influence Maximization using an Infection approach using a graphical probabilistic model. They also
Model from Epidemiology and Topic Modeling to derive a deal with the effciency problem by devising a distributed
unified social score for each node in the network. learning algorithm under the Mapreduce programming model.
Moreover, they also discuss the applicability of their ap-
2. RELATED WORK proach to the expert finding problem.
Dynamic processes or flows that occur in a social network
can be classified as either conservative or non-conservative 3. CONTRIBUTION
[10][11].A flow is said to be conservative if the initial mass or Our unique contribution (to the best of our knowledge) to
content of the network is equal to the final mass or content the problem of assigning influence score to the nodes in a
after the flow has taken place. Conversely, a flow is consid- social network based on their online activity (Twitter activ-
ered non-conservative if the final mass of the network does ity) is integrating Pagerank with the overall ”social presence”
not equal its initial mass after the flow has taken place. of the node over a sample period of time. Here the ”social
Using this classification, the Twitter dataset can be divided presence” of a node depends on the collective probabilities
into conservative and non-conservative components. For the of each of his follower spreading his tweet. We can picture
purpose of this paper, we assume that the social graph of this in the framework of a continuous infection model where
users in the network does not change across time and hence the social presence can be understood as how infected the
the flow of influence across the social graph is conservative. node’s follower network is to spread his message.
To evaluate influence of users across the social graph, we
employ conservative models such as PageRank [12] and In- 4. DATASET DESCRIPTION
betweenness Centrality [13]. We use a snowball sample from the micro-blogging social
Since information flow is essentially non-conservative, for network Twitter collected between 2008 and 2009 to ana-
user activity and tweet propagation, we employ a non- lyze user influence and assign influence scores. The dataset
conservative model called influence maximization based on comprises of the following: The dataset is more fully char-
infection models borrowed from the field of epidemiology
[14]. Influence maximization has been studied in the con-
text of viral marketing. The idea of viral marketing is to Table 1: Summary of Social graph and tweets
target the most influential users in the network so as to ac- Attribute Value
tivate a chain reaction driven by word of mouth, in such a Number of nodes 465,107
way that with a very small marketing cost, we can actually Number of edges 836,541
reach a very large portion of the network. Number of nodes 385,547
The problem of influence maximization is as follows: given Number of tweets 10.5 million
a network with influence estimates, how to select an initial
set of k users such that they eventually select the largest
acterized in [7][8]. Here we provide only the relevant details
number of users in the network. Domingo and Richardson
of the dataset. It has 3 components:
[15] first studied the problem of propagation of influence
and the selection of influential users to maximize impact.
They offered a probabilistic model of interaction and pro- 1. The social graph.
vided heuristics to chose influential users. Kempe et al [14]
tackled the same problem from the perspective of discreet 2. The User Profile information.
optimization . Their work focuses on two fundamental prop-
3. The Tweets data.
agation models, namely Linear Threshold Model and Inde-
pendent Cascade Model.
Leskovec et al [16] study the propagation problem from a The social graph has 2,502 seed nodes that are connected
different perspective namely outbreak detection: how to se- to a total of 465,107 followers. The user profile information
lect nodes in a network in order to detect the spread of a captures the most basic aspects of the user such as its time
virus as fast as possible? They present a general methodol- zone, number of followers, number of people the user is fol-
ogy for near optimal sensor placement in these and related lowing etc. The Tweet data comprises of 10.5 million tweets
problems. By exploiting submodularity they develop an effi- from approximately 118,000 users that are listed in the so-
cient algorithm based on a”lazy-forward” optimization in se- cial graph and the user profile data.
lecting new seeds, achieving near optimal placements, while We also searched the Tweet dataset and counted the number
being 700 times faster than the simple greedy algorithm. In of messages that have been retweeted once, twice and so on.
spite of this big improvement over the basic greedy algo- The table below describes the frequencies of tweet diffusion
rithm, their method still faces serious scalability problems which follows a pareto distribution:
as shown in [17].
In that paper, Chen et al. improve the effciency of the
greedy algorithm and propose new degree discount heuris-
5. PROPOSED METHODOLOGY
tics that produce influence spread close to that of the greedy 5.1 Influence score
algorithm but much more effciently. Tang et al. [18] in-
troduce the problem of topic-based social influence analysis. To assign an influence score to the users in Twitter let us first
Given a social network and a topic distribution for each user, describe the random surfer model and then design a twitter
the problem is to find topic-specific subnetworks, and topic- analog to it. A websurfer browses the web by randomly
specific influence weights between members of the subnet- folowing the hyperlink structure. For example he will arrive
# of times a tweet is retweeted Count let NU be the set of followers of U and OU be the set of
1 727902 followees of U , and we define the following:
2 51980 X
3 4926 Inf luence(U ) = pU (V )IU (V )R(V ). (3)
4 692 V ∈NU

5 188 For simplicity we will assume that pu V is uniform and equal


6 65 to kO1V k . If some twitterers follow one another in a looping
7 30 manner and thereby increasing each others score, we can
8 11 tackle this issue by adding a teleportation vector, similar to
9 5 the random surfer model
10 4
1−d X R(V )
11 2 Inf luence(U ) = IU (V ) . (4)
n V ∈N kOV k
U
Table : Tweet diffusion table of values
An interesting question is whether it is possible to identify
the users who are good at spreading information. In Twit-
at a page and choose randomly one of the outlinks that ter a good measure of influence of an user B on A is how
lead him to a new page and continue. In a long run the many times A has retweeted B. It is not clear how to use
importance of a page is measured by the proportion of time the pairwise influence measures to assign a relative influence
spent on a page. Typically a page that is linked by many score that each user has on the whole network. In this model
others will be important as the surfer will revisit this page we will assign a score to each user which will determine how
often. There are two problems he may encounter. He may likely he is to retweet after reading a tweet. So we will deter-
be stuck in a loop, for example Page 1 → Page 2 → Page 3 mine I(V ), independent of the followee, for each user based
→ Page 1, assuming that the pages have no other outlinks on a how effective the node is for information diffusion. The
than the ones mentioned. Or he lands on a page that has no model used for this purpose is the one used to model the
outlink (dangling page). To avoid getting stuck in the above spread of infectious disease.
situations a surfer should have the option of jumping out by
teleporting to a page which is not in the path. Defining the 5.2 Infection driven score
following
In a network defined by a set of nodes and edges connecting
• Rank(U ): Rank of a page U . them, each node can be 2 or more states. Assuming that
each node can interact with any other node the infection
• NU : The set of all pages pointing into page U (inlinks model describes the interaction using a system of coupled
to U ). differential equations. In this model ( SI model) each node
can be in two states, S for susceptible and I for infected. In
• OU : The set of all pages pointing out from page U . this model for each node the rate of change of the likelihood
of catching the disease is based on the local influence. Let
Ii (t) be the probability that a node is in state I and assume
The basic page ranking algorithm assigns the following rank that the node i in state S and node j in state I are inde-
to a page U pendent events. The adjacency matrix Aij determines the
X R(V ) edges between the nodes where,
R(U ) = (1)
V ∈N
kOV k Aij = 1 if j follows i (5)
U
= 0 else.
where kOV k is the number of elements in the set OU . The
probability that a follower V will read a tweet by U will Then the following differential equation determines the SI
depend on the number of followees of V . In random surfer model,
model a damping factor d is incorporated which directs the
surfer to follow the link d% of the time and teleport to a dIi (t) X
= β(1 − Ii (t)) Aij Ij (t), (6)
new page (1 − d)% of the time. If the total number of web dt j
page is N , I(0) = I0
1−d X R(V )
R(U ) = +d . (2) where β is the transition rate. The nonlinear term in the
N V ∈N
kOV k right hand side makes it difficult to study the differential
U
equation. We wishP to study how Ii (t) evolves for small time
Let us understand how the random surfer model can be de- and we will ignore Aij Ij (t)Ii (t). The above model reduces
rived in Twitter scenario, where we will measure the influ- j
ence of an user by the number of people who will read his to
tweets. Let us study the possible actions after an member dI(t)
U in Twitter tweets. A follower V of U will read the tweet = AI, (7)
dt
with probablity pU (V ) and then retweet with a probability I(0) = I0
IU (V ). The influence of the U will also depend on the influ-
ence of his followers which effects the number of people who where I(t) = (I1 (t), I2 (t), ..., In (t))t is vector determining
will read the tweet by U . Similar to the above notation the ”infection” of each node. The initial value problem (7)
has a closed form solution. To make the assumption of ig-
noring the quadratic term in (6) reasonable, we can start Table 3: Top 10 ”Infected”members at time t = 0
with the initial condition I0 small (initially the network is User Tw Fout Fin I(0)
not infected) and let it evolve according to the linear model ’2020plus1’ 168 807 4 0.1635
for a short period of time. The following lemma proves that ’3sistersmkt’ 199 500 0 0.1200
the infection will not change unreasonably in a short time. ’abpagfh’ 200 496 0 0.1196
’cell49’ 200 491 0 0.1184
’3boyzmom’ 198 491 0 0.11724
Lemma: Given a time interval [t1 , t2 ], the change of infec- ’2ne1one’ 197 493 0 0.11713
tion I(t) over this time interval according to the initial value ’bretrowe’ 199 488 2 0.11711
problem (7) satisfies ’roberawards’ 198 490 0 0.1170
’estevanvanzant’ 200 484 1 0.1167
|I(t1 ) − I(t2 )| ≤ βM |t1 − t2 | (8) ’123print’ 196 493 2 0.1165
where M is a constant.
Proof: Let (λ1 , λ2 , ..., λk ) and (v1 , v2 , ..., vk ) be the eigenval-
ues and eigenvectors of matrix A. Then any vector I(t) in Table 4: Top 10 ”Infected”members at time t = 10
Rn can be expressed in terms of the eigenvectors as User Tw Fout Fin I(10)
’2020plus1’ 168 807 4 0.1808
’a2dsportstalk’ 196 487 5 0.1587
X
I(t) = wl (t)vl , (9)
l ’30wastedyears’ 180 500 4 0.1319
’aboslovesyou’ 200 474 2 0.1299
where the weights wl depends on time. Substituting the ’123 marketing’ 173 451 4 0.1290
above form for I(t) in (7), we find that the weights satisfy ’greennoblis’ 200 475 2 0.1290
the ordinary differential equation ’7meditations’ 198 468 3 0.1287
dwl (t) ’3boyzmom’ 198 491 1 0.1277
= βA (wl (t)vl ) = βλl wl (t), (10) ’aanetwork’ 169 496 5 0.1267
dt
wl (0) = wl . ’bretrowe’ 199 488 2 0.1259

Then the solution of the above linear differential equations


with constant coefficients can be written as where ∆t is the time step. The initial vector I(0) is assigned
in terms of the number of tweets and retweets of an user
wl (t) = wl eβλl t . (11)
of tweets by user j
Ij (0) = (1 − α)
Further wl can be expressed in terms of the initial condition total # of tweets by all users
I0 as # of retweets by user j
+ α
hI0 , vl i hI0 , vl i total # of retweets by all users
wl = = . (12)
hvl , vl i kvl k The parameter α determines the effect of tweet vs retweet on
the infection of an user. We study the evolution of the twit-
Therefore the solution I(t) can be written as
ter activity encapsulated by I0 according to the dynamical
X hI0 , vl i βλ t system (7).
I(t) = e l vl . (13)
kvl k2
l
6. RESULTS
Now we can bound the change in I(t) over the interval [t1 , t2 ] We use tables 2, 3, 4 and 5 to summarise the results of the
by using Mean value theorem and Cauchy-Schwartz as ”infection” score. In the tables the first column represents
the member u in the , the second column TW gives the num-
X |hI0 , vl i| βλ t∗ |t −t |
kI(t1 ) − I(t2 )k ≤ e l 2 1 kvl k ber of tweets by u, the third column Fout is the number of
kvl k2 followers of u in the social graph, Fin is the number of fol-
l
X kI0 kkvl k βλ t∗ lowers of u in the ingraph and the last column gives the
≤ |t2 − t1 | e l kvl k infection score I(t) at time step t. The first table repre-
kvl k2
l sents the top 10 infected members calculated according to
the initial condition. It is clear that intially the infection
where t∗ is a point in [t1 , t2 ]. Let λ1 be the principal eigen- score of a member is not amplified by their connections.
value then When we run our model (14) after 10 time steps we see the

kI(t1 ) − I(t2 )k ≤ keβλ1 t kI0 k|t2 − t1 | effect of the twitter activity as well as the connections on
the score. It can be easily verified by studying the fourth
This lemma proves that I(t) is Lipchitz continuous. column in table 2, where none of the members have 0 fol-
lowers in the ingraph. We can also observe that the I value
of ’a2dsportstalk’ increased at a faster rate than ’2020plus1’
To solve the problem numericall we use the forward differ- in the time interval [0, 20]. We can attribute this to the
ence approximating the derivative as number of followers of ’a2dsportstalk’ in the ingraph and
outgraph and hence ’a2dsportstalk’ feels a higher mean field
I(t + ∆t) = I(t) + ∆tAI(t) (14) or averaged effect from its subgraph connection. We can ob-
Now we would like to compare our results for the pager-
Table 5: Top 10 ”Infected” members at time t = 30 ank versus the modified pagerank. To compute the pager-
User Tw Fout Fin I(30) ank algorithm we consider the teleporation parameter d =
’a2dsportstalk’ 196 487 5 0.2638 0.85 and a constant probability that a member will retweet
’2020plus1’ 168 807 4 0.2123 IU (V ) = 0.2. Our algorithm converges fast and we get the
’aanetwork’ 169 496 5 0.1967 following result. We observe that the score is partial to the
’123 marketing’ 173 451 4 0.1952 number of followers and does not demonstrate any depen-
’30wastedyears’ 180 500 4 0.1895 dency on the twitter activity.
’2frog’ 189 440 3 0.1795
’2abetteru’ 192 435 4 0.1763
’7meditations’ 198 468 3 0.1685 Table 8: Top 10 Pagerank scores
’aboslovesyou’ 200 474 2 0.1644 User Tw Fu Pr
’aby’ 157 426 4 0.1587 ’abendkleider’, 43 8 1.1052
’959klzx’ 173 7 1.0202
’abintracenter’ 13 6 0.8537
Table 6: Top 10 ”Infected” members at time t = 60 ’32auctions’ 45 5 0.7652
User Tw Fout Fin I(60) ’aanetwork’ 169 5 0.7369
’a2dsportstalk’ 196 487 5 0.5103 ’3actsofmurder’ 62 9 0.6584
’aanetwork’ 169 496 5 0.3764 ’4sitemarketing’ 128 5 0.6519
’2dbarcodenow’ 89 434 4 0.3271 ’140cinema’ 16 5 0.5527
’30wastedyears’ 180 500 4 0.3232 ’aaronparsons’ 83 3 0.5102
’2frog’ 189 440 3 0.3121 ’4lxvfm’, 1 3 0.5102
’123 marketing’ 173 451 4 0.3063
’3actsofmurder’ 62 489 9 0.2993 Now we implement the modified pagerank algorithm with
’abeautyoflife’ 41 490 3 0.2893 the retweeting probability as I(60). We can immediately
’aby’ 157 426 4 0.2819 see the effect of the influence score on the ’aanetwork’ with
’2abetteru’ 192 435 4 0.2763 highest modified pagerank. The members ’3actsofmurder’,
’123 marketing’ have improved their modified pagerank due
to the infection score. The highest infected member ’a2dsportstalk’
is a follower of ’30wastedyears’ and has boosted his pager-
serve similar behaviour at time step 30, where the members ank.
with higher number of followers are in the top 10 infection
scores. In table 5 we can note that the member ’2dbarcode-
now’ has higher infection score than ’30wastedyears’ inspite Table 9: Top 10 Modified Pagerank scores
of tweeting less and having less number of followers. Upon User Tw Fu Pr
further investigation we found that the followers of ’2dbarco- ’aanetwork’, 169 5 1.3865
denow’ are more susceptible to infection (has more number ’959klzx’ 173 7 1.1623
of followers and tweets) than than the followers of ’30wast- ’3actsofmurder’ 62 9 0.9852
edyears’ (has only one influential follower). By studying ta- ’32auctions’ 45 5 0.7354
ble 4 we can try to predict which member will go up the list ’140cinema’ 16 5 0.6755
in the following time step. Our first candidate is ’3actsof- ’123 marketing’ 173 5 0.5234
murder’ who has 9 followers! Our prediction is correct and ’365movies’ 173 2 0.5078
’30wastedyears’ 180 4 0.5329
’1800hotwater’ 53 2 0.4948
Table 7: Top 10 ”Infected” members at time t = 70 ’773info’ 40 4 0.4774
User Tw Fout Fin I(60)
’a2dsportstalk’ 196 487 5 0.6344
’aanetwork’ 169 496 5 0.4695 7. CONCLUSIONS AND LESSONS LEARNT
’2dbarcodenow’ 89 434 4 0.4092 In designing this model on a theoretical aspect we realised
’30wastedyears’ 180 500 4 0.3232 the simplicity yet robustness of the pagerank algorithm. Af-
’abeautyoflife’ 41 490 3 0.3807 ter a thorough review of the current research and available
’2frog’ 189 440 3 0.3777 literature on influence in social network, we hypothesized
’3actsofmurder’ 62 489 9 0.3729 the novel approach of calculating the overall “infectibility” of
’123 marketing’ 173 451 4 0.3520 the follower graph of each node based on an infection spread
’aby’ 157 426 4 0.3402 model. We incorporated this infection spread model to form
a modified pangerank . The results satisfactorily matched
’140cinema’ 16 471 5 0.5527
our expectation that members with high “infectibility” score
(susceptible to influence) should boost their own pagerank
we observe that ’abeautyoflife’ is in the 5th position now as well as those whom they follow.
as ’a2dsportstalk’ is a follower. We can also observe that On a practical aspect we used Python to do the coding
our proposed model seem to have fixed the top 3-4 infected and learnt that using inbuilt functions does not always pro-
members and slowly reaching the steady state. duce correct result. For example the pagerank algorithm
from networkx library assigns high values to members in
the graph who do not have any followers. The data pro- 9. REFERENCES
cessing step required extensive python coding to clean and [1] Grewal, R., T.W. Cline, and A. Davies, 2003.
characterize the data before implementing the model. Early-Entrant Advantage, Word-of-Mouth
Communication, Brand Similarity, and the Consumer
8. FUTURE GOALS Decision-Making Process. Journal of Consumer
Going forward we would like to implement our model on a Psychology, 13(3).
richer dataset, that has several degrees of social graph con- [2] Gilad Lotan, Erhardt Graeff, Mike Ananny et al., 2011.
nectivity. We would also lke to incorporate “topic modeling” The Revolutions Were Tweeted: Information Flows
into the calculation of influence score according to the for- During the 2011 Tunisian and Egyptian
mula Iu = f (T, u), where T is the given topic and Iu is the Revolutions.International Journal of Communication 5
influence of user u. For this we intend to use topic extrac- (2011), Feature 1375 -1405
tion algorithms such as Latent Dirichlet Allocation (LDA) [3] D.E. Maskim Tsvetovat & A. Kouznetsov, 2011. Social
and Latent Semantic Indexing (LSI) to calculate the topic Network Analysis for Startups. O’Reilly Publication.
affinities of each user from their tweet sets. Pg15-18
[4] Lee Rainie, Paul Hitlin, Mark Jurkowitz, et al. March
15, 2012. The Viral Kony 2012 Video.Pew Research
Center.
[5] The Kony 2012 Campiagn:
http://www.kony2012.com/
[6] Why ’Kony 2012’ video grabbed 100 million
views online, March 16, 2011, Christian Science
Monitor
[7] Munmun De Choudhury, Hari Sundaram, Ajita John,
Doree Duncan Seligmann, Aisling Kelliher, June 2010.
”Birds of a Feather”: Does User Homophily Impact
Information Diffusion in Social Media?
[8] Choudhury, M. D., Lin, Y-R., Sundaram, H., Candan,
K. S., Xie, L., Kelliher, A. How Does the Sampling
Strategy Impact the Discovery of Information Diffusion
in Social Media? 4th Int’l AAAI Conference on
Weblogs and Social Media (George Washington
University, Washington, DC, May 23-26, 2010).
ICWSM 2010.
[9] Twitter Dataset:
http : //web.archive.org/web/20100326073258/http :
//www.public.asu.edu/ mdechoud/datasets.html
[10] S. P. Borgatti. Centrality and network flow. Social
Networks, 27:55-71, 2005
[11] Rumi Ghosh, Kristina Lerman, Predicting influential
users in online social networks (2010). SNA-KDD:
Proceedings of KDD Workshop on Social Network
Analysis, 2010
[12] Lawrence Page , Sergey Brin , Rajeev Motwani , Terry
Winograd. 1999. The PageRank Citation Ranking:
Bringing Order to the Web. Technical Report. Stanford
Digital Library Technologies Project, 1998.
[13] Wikipedia, ’Betweenness Centrality’, http :
//en.wikipedia.org/wiki/Centrality#Betweenness −
centrality
[14] D. Kempe, J.M. Klienbergg, and E. Tardos.
Maximizing the spread of influence through a social
network. In Proc. Of the Ninth ACM Intl. Conf. on
Knowledge Discovery and Data Mining (KDD’03)
[15] P. Domingo & M. Richardson. Mining the network
value of customers. In Proc. Of the 7th ACM Int. Conf.
on Knowledge Discovery and Datamining (KDD,’01).
[16] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J.
VanBriesen, and N. S. Glance. Cost-effective outbreak
detection in networks. In Proc. of the 13th ACM Int.
Conf. on Knowledge Discovery and Data Mining
(KDD’07)
[17] W. Chen, Y. Wang, and S. Yang. Efficient inı̈ňĆuence
maximization in social networks. In Proc. of the 15th
ACM Int. Conf. on Knowledge Discovery and Data
Mining (KDD’09).
[18] J. Tang, J. Sun, C. Wang, and Z. Yang. Social influence
analysis in large-scale networks. In Proc. of the 15th ACM
SIGKDD Int. Conf. on Knowledge Discovery and Data Min-
ing (KDD’09).

Das könnte Ihnen auch gefallen