Misq 42 1 63

RESEARCH ARTICLE
TOP PERSUADER PREDICTION FOR SOCIAL NETWORKS1

Xiao Fang
Lerner College of Business and Economics, University of Delaware,
Newark, DE 19716-3130 U.S.A. {[email protected]}
Paul Jen-Hwa Hu
David Eccles School of Business, University of Utah,
Salt Lake City, UT 84112 U.S.A. {[email protected]}
Top persuaders in a social network are social entities whose adoption of a product or service will result in the
largest numbers of other entities in the network adopting the same product or service. Predicting top
persuaders is critical to an expanding array of important social network-centric applications, such as viral
marketing, customer retention, and political message promotion. This study formulates the top persuader
prediction problem and develops a novel method to predict top persuaders. Our method development is rooted
in eminent social network theories that reveal several forces central to social persuasion, including social
influence, entity similarity, and structural equivalence. Our method innovatively integrates these forces to pre-
dict top persuaders in a social network, in contrast to existing methods that primarily focus on social influence.
Specifically, we introduce persuasion probability that denotes the likelihood of persuasion conditioned on these
forces. We then propose how to estimate persuasion probabilities, develop an algorithm to predict top per-
suaders using the estimated persuasion probabilities, and analyze the theoretical property of the algorithm.
We conduct an extensive evaluation with real-world social network data and show that our method substan-
tially outperforms prevalent methods from representative previous research and salient industry practices.
Keywords: Top persuader, influential user, machine learning, social network, social persuasion, social
influence, entity similarity, structural equivalence, data-driven method
Introduction1 social networks is social persuasion that refers to the prin-

ciples and processes by which a social entity’s attitude, belief,
A social network consists of a set of social entities and the or behavior is affected by other entities in a social network
relationships among them (Wasserman and Faust 1994). Such (Chaiken et al. 2000).2 This phenomenon gives organizations
networks range from traditional social networks, such as mail the opportunity to predict essential social entities, namely top
correspondence among people (Travers and Milgram 1969), persuaders, whose adoption of a product, service, or opinion
to technology-enabled social networks, such as those sup- will result in the largest numbers of other entities in a social
ported by social media (e.g., Twitter) or online games (Hemp
2006; Kleinberg 2008). An important phenomenon within
2
Social persuasion closely relates to social contagion (Burt 1987) and net-
work diffusion (Valente 1995), in that these concepts all center on the spread
1
Kartik Hosanagar was the accepting senior editor for this paper. Zhengrui of attitudes, opinions, or behaviors within a social network, with a focus on
Jiang served as the associate editor. potential changes in the attitude or behavior of a person who has not yet
adopted, as a result of his or her exposure to and interaction with others who
The appendices for this paper are located in the “Online Supplements” have adopted the focal product or service (Petty and Cacioppo 1986; Simon
section of the MIS Quarterly’s website (http://www.misq.org). 1976).
DOI: 10.25300/MISQ/2018/13211 MIS Quarterly Vol. 42 No. 1, pp. 63-82/March 2018 63

Fang & Hu/Top Persuader Prediction for Social Networks
network to adopt the same product, service, or opinion. Parti- adoption of a product or service, partly due to the similarity
cularly, in a technology-enabled social network, massive and in their individual characteristics. Furthermore, similarity
rich data reflecting individual social entities’ behaviors are also exists in structural position within a social network.
readily available for developing data-driven methods to Structurally equivalent people occupy the same position in a
predict top persuaders. social structure; specifically, social entities are considered
structurally equivalent if they relate in the same ways to other
Predicting top persuaders is critical to an expanding array of entities in a social network (Lorrain and White 1971). Struc-
important social network-centric applications. For example, tural equivalence matters because the similarity in structural
seeding has emerged as a key strategy for firms to market position within a social network can affect individual entities’
products and services (e.g., viral marketing), for health agen- beliefs, attitudes, and behaviors (Burt 1987). In this vein,
cies to foster healthy lifestyles, and for politicians to promote predicting top persuaders requires proper consideration of
political views or public policies (Carr 2008; Watts and different essential forces underlying social persuasion, a
Peretti 2007). In general, seeding targets a group of social fundamental challenge that motivates this research.
entities as “seeds” and exploits their persuasion effects to
reach out and affect a broader population. The effectiveness To address this challenge, we propose a novel method to
of seeding apparently depends on the persuasion power of the predict top persuaders. Different from existing methods that
selected social entities, which underscores the significance of primarily focus on social influence, our method predicts top
top persuader prediction. Firms can also leverage these per- persuaders by simultaneously considering all three important
suaders for customer retention. According to a Forbes forces underlying social persuasion. In particular, we innova-
Insights report (2011), customer retention has become a top tively integrate these forces by introducing persuasion proba-
marketing priority for firms and often accounts for the largest bility, or the probability that one social entity persuades
portion of their marketing budget. By retaining top another to adopt, given the impacts of social influence, entity
persuaders, a firm not only keeps these customers (i.e., top similarity, and structural equivalence from the former to the
persuaders) but also benefits from their persuasion effects to latter. We then propose how to estimate persuasion proba-
retain other customers they affect. On the other hand, when bilities and how to predict top persuaders with the estimated
these persuaders churn, the firm loses them and other persuasion probabilities. The remainder of the paper proceeds
customers they affect. Therefore, a cost-effective customer as follows. We first review relevant previous research, high-
retention strategy is to focus on retaining top persuaders, light the research gaps to be addressed, and discuss the
which requires accurate prediction of such persuaders. differences between our proposed method and existing
methods. Next, we describe the foundation of our method
Because of its importance, top persuader prediction has development, anchored in social network theories. We then
attracted attention from both academic researchers and formulate the top persuader prediction problem and propose
industry practitioners. In industry, methods built on degree a data-driven method to predict top persuaders. To illustrate
centrality have been developed to identify top persuaders. For its practical value, we empirically evaluate the effectiveness
example, according to Weng et al. (2010), Twitter considers of our proposed method with real-world social network data,
its users with the largest numbers of followers as top persua- using representative existing methods as benchmarks.
ders. In academia, prior research has developed methods for Finally, we summarize our contributions, discuss important
identifying influential individual entities in a social network implications, and conclude with several future research
(e.g., Cha et al. 2010; Ghosh and Lerman 2010). A review of directions.
industry practices and academic research on top persuader
prediction suggests a predominant focus on social influence,
which refers to the process by which a social entity’s attitude,
belief, or behavior is affected by (direct) communications and
Related Work
interactions with his or her neighboring entities (Friedkin
1998; Rice et al. 1990). However, according to social net- Existing methods for identifying top persuaders in a social
work theories, social persuasion can arise from multiple network include centrality-based methods, PageRank-based
distinct forces that include social influence, entity similarity, methods, and TwitterRank. In general, centrality-based
and structural equivalence (Burt 1987; Knoke 1990). Entity methods rank social entities according to a centrality measure
similarity denotes the degree to which two social entities are (Wasserman and Faust 1994) and consider those entities with
“similar” in their individual characteristics. One entity in a high centrality scores influential and thus top persuaders
social network can use other similar entities as “a frame of (Borgatti et al. 2009; Brass 1984; Cha et al. 2010; Ghosh and
reference” and adjust his or her attitude, belief, or behavior Lerman 2010). Since the initial conceptualization of cen-
accordingly. Thus, a social entity could affect another’s trality by Bavelas (1948), different centrality measures have
64 MIS Quarterly Vol. 42 No. 1/March 2018

been developed; among them, degree centrality (Albert et al. comments, number of outlinks, and length, and thereby
2000; Walter et al. 1996), topological centrality (Borgatti et identify top bloggers as those who write influential blogs. For
al. 2009; Brass 1984;), and eigencentrality (Ballester et al. Internet-enabled social networking, Trusov et al. (2010)
2006; Bonacich 1972) are widely applied. The degree of a examine the log-in activities in a social network and consider
social entity refers to the number of other entities to which it influential users as those whose activity level significantly
is connected directly (Freeman 1979). An entity that has a affects others’ activity levels. Romero et al. (2010) argue that
higher degree is likely to communicate and interact with more the influence of a Twitter user depends on not only the
entities in a social network and therefore is more influential. number of his or her followers but also his or her ability to
Although commonly utilized, degree centrality is not particu- overcome the passivity of these followers. They therefore
larly effective for identifying influential entities in a social develop an algorithm to compute the influence and passivity
network (Watts and Dodds 2007). Compared with degree score of each Twitter user, on the basis of the interaction
centrality, topological centrality is more computationally between influence and passivity. Zhao et al. (2014) identify
expensive. A representative topological centrality is close- influential users in an online health community and define an
ness centrality. The closeness of a social entity is calculated influential response as one whose sentiment is aligned with
as the inverse of the sum of its distances to all other entities the sentiment change of a community member; accordingly,
in a social network (Freeman 1979). Clearly, the closer a influential users are those who post the greatest number of
social entity is to other entities in a social network, the more influential responses. Finally, Probst et al. (2013) provide a
efficiently that entity can communicate and interact with detailed survey of top persuader identification from the lens
others and, as a result, the more influential it becomes. of social influence (i.e., identifying influential users).
Another common topological centrality is betweenness cen-
trality. In essence, the betweenness of a social entity indi- While existing methods focus on social influence, entity
cates its frequency of falling on the shortest paths that link similarity could contribute to social persuasion too. Several
pairs of other entities in a social network (Freeman 1979; studies have shown empirically the important role of entity
White and Borgatti 1994). Betweenness reflects a social similarity in driving social persuasion (Anagnostopoulos et al.
entity’s potential of controlling the communications and inter- 2008; Crandall et al. 2008). For example, Crandall et al.
actions between pairs of other entities; the higher an entity’s (2008) examine the relative importance of social influence
betweenness, the more influential that entity becomes. In and entity similarity for predicting social persuasion, and
addition to degree and topological centrality, eigencentrality report that, while both contribute to social persuasion, social
represents another popular centrality measure. Eigenvector influence seems more powerful for predicting social persua-
centrality, an important eigencentrality, weights directly con- sion on Wikipedia whereas entity similarity appears more
nected social entities according to their centrality scores important for predicting social persuasion on LiveJournal.
(Bonacich 1972, 1987, 2007). Another important eigencen- Although prior literature has recognized the significance of
trality is intercentrality, which considers both a social entity’s entity similarity, few studies consider both entity similarity
centrality and its contribution to every other entity’s centrality and social influence to identify top persuaders in a social net-
(Ballester et al. 2006). Intercentrality can be applied to iden- work. One exception is Weng et al. (2010), who develop a
tify the key player in a network of delinquents whose removal method, namely TwitterRank, to identify influential Twitter
would lead to the highest aggregate delinquency reduction users on the basis of the tweet topic similarity (i.e., entity
(Ballester et al. 2010). Different from centrality metrics similarity) and the following relationship (i.e., social influ-
described above, percolation centrality measures the centrality ence) between Twitter users.
of a social entity according to its percolation state, in addition
to its topological connectivity (Piraveenan et al. 2013). According to our literature review, most existing methods
identify top persuaders from the lens of social influence,
Previous research also analyzes weights of relationships from except for TwitterRank, which combines social influence and
the lens of social influence and incorporates such weights to entity similarity. However, social network theories suggest
identify top persuaders, in general or in a particular applica- that multiple distinct forces—namely, social influence, entity
tion domain. One prevalent method uses the PageRank similarity, and structural equivalence—jointly determine
algorithm (Brin and Page 1998) to rank individual entities social persuasion (Burt 1987; Knoke 1990). Anchored in
according to their social influences (Romero et al. 2010). eminent social network theories, our proposed method differs
Several other studies develop methods to identify top per- from existing methods in that it simultaneously considers
suaders in a particular application domain. Taking blogging these distinct forces crucial to social persuasion. Further-
as an example, Agarwal et al. (2008) measure the influence of more, our method integrates these forces to predict top per-
a blog, using such properties as number of inlinks, number of suaders by introducing persuasion probability that denotes the
MIS Quarterly Vol. 42 No. 1/March 2018 65

Table 1. Comparison of Our Method and Existing Methods

Centrality-Based PageRank-
Methods Based Methods Twitter Rank
(e.g., Ghosh and (e.g., Romero et (Weng et al.
Our Method Lerman 2010) al. 2010) 2010)
Considers social influence? Yes Yes Yes Yes
Considers entity similarity? Yes No No Yes
Considers structural equivalence? Yes No No No
Integrates forces underlying social persuasion Yes No No No
through persuasion probability?
Notes: A method that considers structural equivalence would take into account the persuasion effect between structurally similar social entities
in a social network.
likelihood of persuasion conditioned on these forces. Speci- echo the importance of the linkages that connect individual
fically, we propose how to estimate persuasion probabilities entities in a social network. People tend to compare them-
and then how to use the estimated persuasion probabilities to selves to those with whom they have close ties, then seek to
predict top persuaders. In contrast, Weng et al. (2010) com- evaluate and emulate the attitudes or actions of these
bine social influence and entity similarity to identify top intimates. Interpersonal communications and interactions, via
persuaders by multiplying them without proper theoretical network ties, are recurrent and often prevail in small, close-
justification. Finally, we demonstrate, through an extensive knit neighborhoods within the grand social network. In this
evaluation with real-world social network data, that our
vein, Agarwal et al. (2009) highlight the importance of the
proposed method significantly outperforms representative
existing methods. In Table 1, we compare our method and social influence of local reference groups and geographical
existing methods. peer effects in a social network. Similarly, Susarla et al.
(2012) emphasize the effect of social influence for diffusing
user-generated content in social networks.
Theoretical Foundation Both social comparison theory and social influence network
theory also shed light on the importance of individual
Social influence constitutes a common focus of previous top characteristics. People in a social network may exhibit
persuader research, and its effects can be explained by social similar opinions or behaviors, simply because they share
comparison theory (Festinger 1954). According to this individual characteristics (Aral et al. 2009; Shalizi and
theory, persuasion occurs when a person influences others Thomas 2011). All else being equal, individuals who are
through his or her ability to reduce the uncertainty in a deci- more similar in their demographics, cultural backgrounds, or
sion situation that those others face. Pfeffer et al. (1976)
personal preferences are more likely to have common views
specifically refer to such social influences as informational
social influences. When interacting with others in a social and behaviors. According to attribution theory (Weiner 1974,
network, a person conveys a view that could influence those 1986), people gain cognitive control over their environment
others’ opinions or behaviors by simplifying their evaluation by explaining and understanding the causes of behaviors and
or decision task at hand, such as by reducing the range of environmental occurrences. In this light, people who share
viable alternatives or making a recommendation. To create highly similar characteristics or backgrounds likely have
social influence, the initiator usually sends a message, issues comparable understandings of and reasoning for underlying
an opinion, or delivers a well-articulated view to recipients to causes and therefore exhibit similar opinions and behaviors.
evoke their response (Wood 2000; Zanden 1987).
Structural characteristics can also determine the extent to
Friedkin (1998) further explains the effects of social influence which an entity affects others’ opinions and behaviors (Fang
with social influence network theory, which suggests that et al. 2013; Wejnert 2002). Burt (1987) analyzes social
each entity is endowed with an initial opinion or attitude, contagion for innovation diffusion in social structural circum-
receives and responds to the information disseminated in the stances and suggests that contagion arises when people who
network, and then modifies his or her opinion or attitude are proximate in the structure use one another to manage the
accordingly. The magnitude of the resulting social persuasion uncertainty of an innovation. Thus both cohesion and struc-
depends on the social influence carried through the interentity tural equivalence models are crucial. The structural equiv-
communications and interactions. Such effects and influences alence model emphasizes entities’ structural positions in a

network; the cohesion model focuses on socialization between Problem Formulation

an ego (an entity that has not yet adopted) and an alter (an
entity that has adopted already) in a network. According to Let V denote a set of n social entities vi in a social network
the cohesion model, the more frequent and empathic the (i.e., vi 0 V, i = 1,2,…,n). Social entities in a network are
interaction between an ego and an alter, the more likely an connected by pairwise relationships. A relationship can be
adoption by the latter is to trigger an adoption by the former. directional or nondirectional (Wasserman and Faust 1994).
Burt et al. (1994) also use a contagion model to describe how For example, a social network of avatars in an online game
a network structure draws individuals together, such that consists of a set of avatars (i.e., social entities) connected by
one’s opinion or behavior affects those of the other. Struc- nondirectional communication relationships, whereas a social
tural equivalence can explain observable changes in people’s network on Twitter consists of a set of users (i.e., social
opinions or behaviors in the presence of information entities) connected by directional relationships (e.g., one
dissemination in a social network. Twitter user following another user). For a social entity vi, let
vector fi represent the profile of the entity (i.e., individual
Structural equivalence also represents a crucial characteristic characteristics of the entity); taking an avatar as an example,
of a social network. People occupying identical structural its profile describes the avatar’s characteristics such as gender
positions in a network tend to exhibit more similar opinions and profession. Let sij denote the strength of social inter-
and behaviors than otherwise, partially because structural actions from entity vi 0 V to entity vj 0 V in the relationship
equivalence can reduce the uncertainty associated with an connecting them (e.g., the number of tweets published by vi
evaluation or decision task that people face and therefore can and retweeted by vj). Specifically, sij is 0 if there is no
affect their attitudes and behaviors, even if no social ties relationship from vi to vj. We summarize important notations
connect them directly. As Burt explains, the spread of an used in this paper in Table 2.
opinion or behavior in a social network is contingent, to some
degree, on the way the structure of the network brings entities We formulate the top persuader prediction problem as fol-
together. If they connect to the same group of others through lows: In a social network, we observe at current time a set of
identical links, individuals likely exhibit similar opinions or social entities that has adopted a focal item (e.g., a product,
behaviors, because they vicariously experience or even mimic service, or opinion). The objective is to predict top K
each other, which reduces the uncertainty associated with persuaders (i.e., K individual entities whose adoption of the
their adoption of an opinion or behavior (Rice and Aydin item will result in the largest numbers of other entities in the
1991). network to adopt in a future time period).
Burt emphasizes cohesion and structural equivalence for To solve this problem, we need to properly address two chal-
social contagion that closely relates to the social persuasion lenges. First, for each ordered pair of social entities vi and vj,
we study. According to Knoke (1990), persuasion and selec- we need to estimate the probability that vi persuades vj to
tion processes are fundamental to any comprehensive adopt (i.e., persuasion probability). Second, we have to pre-
explanation of how individuals’ social relationships shape dict top persuaders on the basis of the estimated persuasion
their attitudes and activities. Therefore, persuasion and entity probabilities. In the following, we detail how we address
similarity are both important. Persuasion, as described by each challenge.
Knoke, and cohesion, as defined by Burt, intrinsically cor-
respond to the social influence we examine. By anchoring in
eminent social network theories and integrating the insights
Estimating Persuasion Probabilities
by Burt and Knoke, we consider three forces central to social
persuasion: social influence, entity similarity, and structural
The probability that vi persuades vj to adopt depends on the
equivalence.
social influence from vi to vj as well as the entity similarity
and structural equivalence of vi to vj. We use the strength of
social interactions to measure social influence, because social
Top Persuader Prediction influence results from and therefore is manifested by social
Problem and Method interactions (Rice et al. 1990). Accordingly, the social
influence Iij from vi to vj is measured as
In this section, we formulate the top persuader prediction
sij
problem and then detail our method for predicting top Iij = (1)
persuaders.  s
vh ∈V ,h ≠ j hj

Table 2. Notation
V = set of social entities in a social network vi = social entity, vi 0V
fi = profile of vi sij = strength of social interactions from vi to vj
Iij = social influence from vi to vj Mij = entity similarity of vi to vj
Rij = structural equivalence of vi to vj dij = distance from vi to vj
zij = normalized distance, see Equation (5) γ = attenuation factor, 0 < γ < 1
Dj adoption state of vj pij = persuasion probability, see Equation (8)
ci = persuasion score of entity vi C = vector of persuasion scores
Equation (1) measures social influence from vi to vj as the ( )

ed vi , v j =
proportional social interaction strength from vi to vj, in rela- 1/ 2
  (3)
tion to the total social interaction strength from all vh 0 V to vj,
(
 z −z
)  (z )  (z ) 
2 2 2
+ − z jk + − zkj
h … j. The way we measure social influence is appropriate  ij ji ik ki 
 v k ∈V , k ≠ i , j v k ∈V , k ≠ i , j 
and has been commonly used by previous research (Kempe et
al. 2003; Romero et al. 2010). Social influence Iij is within
where zij is the normalized path distance from vi to vj.
the range of [0,1]; the higher the value of Iij, the more power-
Similarly, ed(vi, vj) in a social network with nondirectional
ful is the influence of vi on vj, relative to the influence of any
relationships is defined as
other social entity on vj.
1/ 2
We measure entity similarity in a similar way; specifically,  
( ) =  ( ) 
2
the entity similarity Mij of vi to vj is ed vi , v j zik − z jk (4)
 
vk ∈V ,k ≠ i , j 
Mij =
(
sim fi , f j )
 sim( f , f )
vh ∈V ,h ≠ j
h j
(2) For a social network with nondirectional relationships, we
have zij = zji, zik = zkj, and zjk = zkj. Hence, the first and third
terms in Equation (3) do not appear in Equation (4).
Following Burt, we define normalized path distance zij as
where sim() is a similarity function that takes the profiles of
two social entities as input and returns their profile similarity.
 dij
The choice of the similarity function depends on the types of  , if there is a path from vi to v j
attributes in fi (Tan et al. 2006). For example, matching zij =  n (5)
coefficient is an appropriate similarity function if fi only con-  1, if there is no path from vi to v j
tains nominal attributes, whereas a heterogeneous similarity
function is proper if fi contains both nominal and numeric where dij denotes the distance from vi to vj and n is the number
attributes. The specific similarity function used in our evalua- of social entities in a social network. As its name implies, the
tion study is discussed in the next section. Entity similarity Euclidean distance function ed() measures structural inequi-
Mij is within the range of [0,1]; the higher the value of Mij, the valence between social entities in a social network. By
more similar vi is to vj, relative to the similarity of any other applying monotonically decreasing transformation, which is
social entity to vj. commonly used for transforming inequivalence to equivalence
(Tan et al. 2006), we define the structural equivalence
Perfect structural equivalence is rare in real-world social function se() as
networks; it is common to measure the structural equivalence
(or inequivalence) between social entities as the extent to
which they are structurally similar (or dissimilar) according to (
se vi , v j = ) 1
( )
(6)
some distance function, such as the Euclidean distance func- 1 + ed vi , v j
tion (Wasserman and Faust 1994). According to Burt (1987),
the Euclidean distance between social entities vi and vj in a
We thus measure the structural equivalence Rij of vi to vj as
social network with directional relationships is defined as

Rij =
( )
se vi , v j adopted. If a social entity vk adopted before current time, we
set its adoption decision Dk = 1 and its adoption decision time
 v ∈V ,h≠ j se(vh , v j )
h
(7)
as its time of adoption. If a social entity vk has not adopted
until current time, we set Dk = 0 and its adoption decision time
as T.3 According to prevalent social diffusion and persuasion
Structural equivalence Rij is within the range of [0,1]; the models in social networks (Granovetter 1978; Kleinberg
higher the value of Rij, the greater the structural equivalence of 2007), a social entity’s adoption decision is affected by other
vi to vj, relative to the structural equivalence of any other social entities that have already adopted, through their impacts of
entity to vj. social influence, entity similarity and structural equivalence.
Therefore, a social entity vk’s adoption decision Dk is affected
Having introduced social influence Iij, entity similarity Mij, and by
structural equivalence Rij from vi to vj, we now define per-
suasion probability pij that vi persuades vj to adopt. According Ik = I
vi ∈VD
ik
to salient social network theories, social persuasion is driven
by social influence, entity similarity, and structural equiva-
lence. We therefore define, in Equation (8), persuasion
Mk = γ
vi ∈VD
dik −1
Mik
probability as the probability that vj adopts, given the impacts

of social influence, entity similarity, and structural equivalence
Rk = γ
vi ∈VD
dik −1
Rik
from vi to vj. That is,
where VD is the set of entities that adopted before vk’s
pij = P D j = 1| Iij ,γ ij Mij ,γ ij Rij 
d −1 d −1
(8) adoption decision time, γ is the attenuation factor, and dik
denotes the distance from vi to vk. We can construct training
where Dj = 1 represents the adoption by vj and Dj = 0 implies data by calculating Ik, Mk, and Rk and setting Dk for each social
otherwise. Furthermore, γ denotes the attenuation factor, entity in a social network. The constructed training data thus
d −1
where 0 < γ < 1, and γ
ij
models the scenario in which the have n records, where n is the number of social entities in the
impact of entity similarity (or structural equivalence) from vi social network. Each record consists of predictor attributes Ik,
to vj attenuates as the path distance dji from vi to vj increases Mk, and Rk and class label attribute Dk .
(Jackson and Wolinsky 1996; Katz 1953; Watts 2001). Fang
et al. (2013) develop an expectation–maximization (EM) based To estimate pij from the training data, a potential method is
method to predict adoption probabilities. The major difference decision tree induction with Laplace estimation, which has
between adoption probability and persuasion probability is the been used to estimate class membership probabilities (Pazzani
consideration of hidden power by the former but not by the et al. 1994). This method first learns a decision tree from the
dij −1
latter. According to Fang et al., adoption probability is training data. According to the value of I ij , γ Mij , and
affected by interentity impacts (including social influence, dij −1
structural equivalence, and entity similarity) as well as γ Rij , a leaf node of the learned decision tree is reached
unobserved factors (i.e., hidden power). Therefore, it is next. The Laplace estimation of pij is then produced by
reasonable to include hidden power in the definition of adop-
tion probability. In this study, persuasion probability is the nD + 1
probability that one social entity persuades another entity to pij = (9)
adopt (i.e., interentity impacts). Thus, our definition of per-
nL + 2
suasion probability focuses on interentity impacts only and
does not consider hidden power. The EM-based method by where nL is the number of training records in the reached leaf
Fang et al. is developed to handle hidden power which, how- node and nD is the number of training records with Dk = 1 in
ever, is not relevant to persuasion probability. Therefore, we the node. Although commonly used, this method has a
propose a novel method to predict persuasion probabilities, limitation. Laplace estimation assumes equal prior proba-
instead of using the EM-based method. bilities, which does not hold in most real-world scenarios
(Zadrozny and Elkan 2001). In our study, Laplace estimation
To estimate pij, we need to construct training data from the assumes that adoption and nonadoption cases are equi-
currently observed adoption information and develop a method probable (i.e., P(Dk = 1) = P(Dk = 0) = 1/2). This assumption
for estimating pij from the training data. At current time T, we
observe the adoption decision of each social entity in a social
3
network and the adoption time for those who have already For a social entity that has not adopted until current time T, its latest
observable adoption decision was made at T.

is unrealistic, because nonadoption cases often substantially depends on the probabilities that vi persuades other entities to
outnumber adoption ones (Aral et al. 2009; Fang et al. 2013). adopt (i.e., pij where vj 0 V and j … i). Understandably, the
Furthermore, prior studies have shown, theoretically and higher such probabilities, the greater vi’s persuasion score ci.
empirically, the use of bagging to significantly improve In addition, persuasion score ci also depends on the persua-
probability estimation accuracy by decision trees (Breiman sion scores of the social entities that vi persuades. Intuitively,
1996). Bagging creates bootstrap samples from the training vi could have a high persuasion score if the entities persuaded
data,4 learns a decision tree from each sample, and yields the by vi have high persuasion scores themselves. We therefore
average of probabilities, each of which is estimated by a calculate ci as
decision tree. Hence, we can use bagging to enhance the
accuracy of estimating pij. ci = 
v j ∈V , j ≠ i
pij c j (11)
We therefore propose to estimate pij using bagged decision tree

induction with m-estimation. Instead of assuming equal prior  c1 
probabilities, m-estimation effectively incorporates prior  
probabilities gauged from the training data for probability  c2 
Let the vector of persuasion scores c =
estimation (Cussens 1993). Previous research has shown that  , each element
m-estimation produces more accurate probability estimations  
than does Laplace estimation (Sulzmann and Fürnkranz 2009).  cn 
Different from Equation (9), the m-estimation of pij is given by of which denotes a social entity’s persuasion score. Let P be
the n × n matrix of persuasion probabilities, where n is the
nD + mP( Dk = 1) number of social entities in a social network, a non-diagonal
pij = (10) element of the matrix pij is the probability that vi persuades vj
nL + m to adopt, and all diagonal elements are set to 0. By Equation
(11), we have
where parameter m can be set as n L (Cussens 1993) and C = PC (12)
prior probability P(Dk = 1) can be estimated as the ratio of
records with Dk = 1 in the training data (Mitchell 1997). As To predict top K persuaders, we need to compute C with
we show in Figure 1, our proposed method consists of two Equation (12); top K persuaders are those with the highest K
phases: training and probability estimation. In the training persuasion scores in the computed C.
phase, a set of decision trees is constructed, each of which is
learned from a bootstrap sample of the training data using In response, we propose an algorithm to compute C and
C4.5, a widely used decision tree learning algorithm (Quinlan predict top K persuaders, based on the power iteration method
1993). Following Breiman (1996), we create 50 bootstrap (Golub and Van Loan 1996). We then show the convergence
samples and build 50 decision trees accordingly.5 In the of the proposed algorithm and specify how to initialize C for
probability estimation phase, pij is estimated as the average of the algorithm (see Figure 2).
probability estimations across 50 decision trees.
The proposed algorithm, shown in Figure 2, first initializes C,
then iteratively updates C according to Equation (12) until
Predicting Top Persuaders convergence, and finally outputs social entities with the
highest K persuasion scores in the converged C. Two details
We define the persuasion score of a social entity as the extent of the algorithm warrant further explanation. First, ||Cx+1||4
to which that entity can persuade other entities to adopt. yields the maximum of the absolute values of Cx+1’s elements:
Accordingly, top K persuaders refer to social entities with the ||Cx+1||4 = max
1≤ i ≤ n
||Cxi+1|; and Cx+1/||Cx+1||4 normalizes Cx+1. This
highest K persuasion scores. For a social entity vi 0 V, let ci $ normalization is a commonly accepted practice for power
0 be the persuasion score of that entity. Persuasion score ci iteration to avoid the risk of overflow or underflow (Heath
2002). Moreover, such normalization does not change
4
A bootstrap sample contains the same number of records as the training data;
persuasion score ranks of individual entities and therefore
each sample record is randomly drawn from the training data with replacement does not affect the final result of predicted top K persuaders.
(Breiman 1996). Second, one-norm ||Z||1 of a vector Z is defined as the sum of
n
5
There is no theory regarding how many bootstrap samples to create for
the absolute values of its elements, that is, ||Z||1 = 
i =1
|Zi |
bagging (Breiman 1996); we follow the rule of thumb of 50 bootstrap samples
suggested by Breiman.

dij − 1 dij − 1
Input: training data Input: I ij , γ M ij , γ Rij
Output: 50 decision trees
Output: pij
Create 50 bootstrap samples from the training data.
Learn a decision tree from each sample using C4.5. For each decision tree learned
Reach a leaf node according to
dij − 1 dij − 1
I ij , γ M ij , γ Rij .
Estimate pij using Equation (10).
End for
Output average pij across 50 decision trees.
(a) Training (b) Probability Estimation
Figure 1. Bagged Decision Tree Induction with M-Estimation
Input: K, P
Output: top K persuaders
Initialize C = C 0.
x = 0.
Do
C x+1 = PC x.
C x+1 = C x+1/||C x+1||4.
μ = ||C x+1 – C x||1/n.
x = x + 1.
Until μ < 0. //0: predefined convergence threshold
Output social entities with the highest K persuasion scores in C x.
Figure 2. Predicting Top K Persuaders
(Heath 2002). Therefore, ||Cx+1 – Cx||1/n represents the entities have same persuasion scores initially.
average absolute difference of persuasion scores between two
consecutive updates of C. The iterative process of our
algorithm terminates if the difference is sufficiently small. Empirical Evaluation and Results
We now show the convergence of the algorithm.
We evaluated our proposed method with real-world social
Theorem 1: The algorithm for predicting top K persuaders network data to demonstrate its superior prediction perfor-
converges and the converged C is unique, independent of the mance over several prevalent methods, and then empirically
initialization of C. analyzed its important properties. In this section, we describe
the data and benchmark methods, detail our evaluation design,
Proof: See Appendix A. and report important evaluation results.
Theorem 1 guarantees the convergence of our algorithm and
assures the uniqueness of the converged C. In addition, Data and Benchmark Methods
Theorem 1 stipulates that both the convergence of the
algorithm and the uniqueness of the converged C are inde- We conducted our evaluation in the context of a social net-
pendent of the initialization of C. Hence, a random initiali- work of avatars. Avatar social networks have been popularly
zation of C is sufficient for the algorithm. Without loss of used to analyze or evaluate social network theories and
 1 methods (Bainbridge 2007; Centola 2010). Specifically, our
 
1 evaluation employed three related data sets collected from a
generality, we initialize C as C 0 =   ; that is, all social
 major online game company. One data set contains 1.34
  million records of complete message communications among
 1 all avatars of an online game over 20 weeks, starting from the

Table 3. Summary of Avatar Profile Attributes

Attribute Name Attribute Type Description
Gender Nominal Avatar gender: male or female
Profession Nominal Avatar profession, such as farmer, solider etc.
Level Ordinal Avatar level in the game, ranging from level 1 (lowest) to level 30 (highest)
Community Nominal Community to which an avatar belongs
Gold Numeric Amount of gold possessed by an avatar
Silver Numeric Amount of silver possessed by an avatar
Bonus Numeric Bonus points earned by an avatar
Experience Numeric Avatar score in experience
Intelligence Numeric Avatar score in intelligence
Strength Numeric Avatar score in strength
Agility Numeric Avatar score in agility
Spirit Numeric Avatar score in spirit
Attack Numeric Avatar score in attack capability
Defense Numeric Avatar score in defense capability
d − mind
launch of the game. Each record consists of the timestamp of
a message communication, and the respective identities of the ( )
sim fik , f jk = 1 −
maxd − mind
two avatars participating in the communication. Using this
data set, we constructed a social network of avatars, in which where difference d = |fik – fjk| and maxd and mind denote the
a social entity represents an avatar and a relationship between maximum and the minimum among all differences respec-
two entities exists if there are message communications tively. The profile similarity sim(fi, fj ) between fi and fj is
between their corresponding avatars. We further measured then calculated as the average attribute similarity across all
the strength of social interactions between two social entities the profile attributes.
using the number of messages communicated between their
corresponding avatars. For each of the 5,162 avatars, we As we describe in our review of related research, salient prior
collected data regarding whether the avatar adopted a methods focus on social influence and often employ a
particular virtual item during the study period and, if adopted, centrality measure to identify top persuaders as those that
in which week. This particular virtual item is not a necessity score high in the measure. We therefore benchmarked our
for avatars. Following extant literature (e.g., Bass 1969; Aral proposed method against existing methods developed on the
et al. 2009), we considered the initial purchase of the item as basis of major centrality measures, including degree
the adoption of that item. Over the study period, a total of centrality, topological centrality, percolation centrality, and
1,021 avatars adopted the virtual item. The third data set con- eigencentrality. According to Freeman (1979), the degree
tains individual profiles of the 5,162 avatars. As we summar- centrality deg(vi) of a social entity vi is measured as
ize in Table 3, the profile of an avatar includes 14 attributes. n
deg(vi ) =  aij
Because of the mixed attribute types in each profile, we j =1
applied a heterogeneous similarity function to compute profile
similarity (Tan et al. 2006). For two avatar profiles fi and fj , where aij = 1 if there exists a relationship that connects
let fik and fjk be their respective kth attribute, k = 1,2,…,14. entities vi and vj, aij = 0 otherwise, and n is the number of
According to Tan et al. (2006), if the kth attribute is nominal, social entities in a social network. Unlike degree centrality,
the attribute similarity sim(fik, fjk) between fik and fjk is topological centrality, such as closeness or betweenness,
given by incurs more computational cost. According to Freeman, the
closeness centrality cls(vi) of entity vi is evaluated as
0 if fik ≠ f jk
( )
sim fik , f jk = 
cls(vi ) =
1
1 if fik = f jk

n
d
j =1 ij
and if the kth attribute is ordinal or numeric, sim(fik, fjk) is
defined as

where dij denotes the distance from entity vi to entity vj; and random selection method. In Table 4, we summarize each
the betweenness centrality btn(vi) of entity vi is given by benchmark method included in our evaluation.
n n
 
In terms of time complexity, computing the degree, closeness,
btn(vi ) = b jk (vi ) and betweenness centrality for each entity in a social network
j = 1 k = 1, k > j takes O(r), O(n3), and O(n3) time respectively (Zhuge and
Zhang 2010), where r is the number of relationships and n
where bjk(vi) = gjk(vi)/gjk, gjk is the number of the shortest paths denotes the number of entities in a social network. Percola-
linking entity vj and entity vk, gjk(vi) is the number of the tion centrality is a variation of betweenness centrality and it
shortest paths linking entity vj and entity vk that contain entity takes O(n3) time to calculate the percolation centrality for
vi, and i … j … k. Unlike betweenness centrality, percolation each entity in a social network. The time required to compute
centrality considers percolation states of social entities, which the intercentrality for each entity in a social network is
are adoption states of entities in our study; and the percolation dominated by the time of multiplying two n × n matrices,
centrality pc(vi) of entity vi is defined as (Piraveenan et al. which takes O(n3) time. For eigenvector centrality, INF-
2013) RANK, INF-SIM, and our method (given persuasion proba-
bilities), the respective running time is dominated by the time
1 n n Dj
pc(vi ) =   bjk (vi ) of multiplying an n × n matrix and an n × 1 vector, which
n − 2 j −1 k =1,k > j
[ D ] − D
n
h =1 h i takes O(n2) time. It is worth noting that our method incurs
additional time to estimate persuasion probabilities. Ac-
cording to our time complexity analysis, the running time of
where Dj = 1 if entity vj adopts, Dj = 0 otherwise, and i … j all the investigated methods, except for degree centrality,
… k. One important eigencentrality metric is eigenvector cen- scales up superlinearly with the number of entities in a social
trality. By representing a social network as an adjacency network. Although degree centrality takes O(r) time, r is
matrix A = [aij]n × n, we can compute the principal eigenvector close to O(n2) for a densely connected social network. There-
of the matrix, which contains eigenvector centrality scores for fore, designing more efficient top-persuader prediction
each social entity in the social network (Bonacich 1972, 1987, methods for large-scale social networks deserves future
2007). Another critical eigencentrality metric is intercen- research attention.
trality. Ballester et al. (2006) define the intercentrality of
entity vi as the square of the number of paths that start at vi
divided by the number of paths from vi to vi itself, where paths Evaluation Design
of length l are weighted by βl, β is a decay factor, and 0 < β
< 1. We first introduce persuasion credit, a concept essential for
evaluating different top persuader prediction methods. The
We also benchmarked our method against a method that persuasion credit of a social entity refers to the credit the
employs the PageRank algorithm to rank social entities based entity receives by persuading other entities to adopt. It is
on their social influences Iij (Brin and Page 1998; Romero et generally accepted to measure the persuasion (or influence)
al. 2010), namely INF-RANK in our evaluation. It is essen- power of a social entity based on other entities’ adoption
tial to compare with prior methods that consider both social behaviors after the entity’s adoption because these subsequent
influence and entity similarity for identifying top persuaders, adoptions provide an observable indicator of an entity’s
though few such methods have been proposed. In particular, persuasion (or influence) power (Bakshy et al. 2011; Trusov
Weng et al. (2010) develop a transition matrix B = [bij]n × n for et al. 2010). We therefore follow this practice to measure
a social network and define bij as persuasion credit. Specifically, persuasion credit is extended
bij = Iij × Mij from influence score, a well-established concept proposed by
Bakshy et al. (2011), which is defined as the score an entity
where social influence Iij and entity similarity Mij are given by obtains by influencing other entities to adopt. We use the
Equations (1) and (2) respectively. The principal eigenvector following example to illustrate influence score and its
of the transition matrix contains rank scores of each social calculation.
entity in the social network. In our evaluation, we refer to this
benchmark method as INF-SIM. It is also important to in- Consider a scenario in which social entity E adopts after the
clude as benchmark a random selection method that randomly adoption by entities A, B, C, and D. According to Bakshy et
selects social entities as top persuaders. The inclusion of this al. (2011), the adoption by entity E can be attributed to the
method provides a comparison baseline; we anticipate our social influence of E’s direct neighbors that have adopted
method and all other benchmark methods to outperform the earlier than E and each direct neighbor receives an equal credit.

Table 4. Summary of Benchmark Methods

Method Force(s) Considered Source(s)
Degree Centrality Social Influence
Freeman 1979; Wasserman and Faust 1994;
Closeness Centrality Social Influence
Borgatti et al. 2009
Betweenness Centrality Social Influence
Percolation Centrality Social Influence Piraveenan et al. 2013
Eigenvector Centrality Social Influence Bonacich 1972, 1987, 2007
Intercentrality Social Influence Ballester et al. 2006
INF-RANK Social Influence Brin and Page 1998; Romero et al. 2010
INF-SIM Social Influence, Entity Similarity Weng et al. 2010
Random Selection / Baseline
Table 5. An Illustration of Influence Score and Persuasion Credit

Social Entity Distance to E Influence Score Persuasion Credit
A 1 0.50 0.38
B 1 0.50 0.38
C 2 0 0.19
D 4 0 0.05
As shown in Table 5, entities A and B are E’s direct the previous section, a social entity’s power of persuading
neighbors because their distance to E is 1. Furthermore, another attenuates as the distance between them increases
entities A and B adopted earlier than entity E. Therefore, the (Jackson and Wolinsky 1996; Katz 1953; Watts 2001). It is
adoption by entity E is attributed to the social influence of A therefore necessary to introduce the attenuation factor 0 < γ
and B and, as a result, each receives half of a credit. As we < 1 in the evaluation of persuasion credit. We continue using
show in Table 5, the respective influence score of A and B is the example in Table 5 to illustrate the calculation of persua-
0.5, because of the contribution of E’s adoption. Entities C sion credit. Let the persuasion credit of a social entity, before
and D are not E’s direct neighbors and hence receive no credit it is discounted by the attenuation factor, be x. Due to the
from E’s adoption. As illustrated in Table 5, the influence adoption by E, we have
score of entity C or D is 0. Table 5 lists influence scores
resulting from E’s adoption only. As more entities become xγ1 – 1 + xγ1 – 1 + xγ2 – 1 + xγ4 – 1 = 1
adopters, additional credits should be added to influence
scores, according to the procedure described above. The four terms in the left hand side of the above equation
represent the respective persuasion credit of entities A, B, C,
Our persuasion credit extends influence score in two ways. and D, after being discounted by the attenuation factor. By
First, influence score measures a social entity’s power of assuming γ = 0.5 and solving the equation, we note that x =
social influence and therefore its consideration of only direct 0.38 and obtain persuasion credits of A, B, C, and D as shown
neighbors of the entity seems reasonable (Bakshy et al. 2011). in Table 5. Table 5 again lists persuasion credits due to E’s
Persuasion credit, on the other hand, measures a social adoption only. As more entities become adopters, additional
entity’s power of social persuasion, which results from the credits should be added to persuasion credits following the
combined effects of social influence, entity similarity, and described procedure.
structural equivalence. The evaluation of a social entity’s
persuasion credit thus requires the proper consideration of Next, we detail our evaluation design. We divided evaluation
both direct neighbors (i.e., those with distance to the entity data into two parts. One part contained data over the first 10
equal 1) and indirect neighbors (i.e., those with distance to the weeks of the study period, which we used to train our method
entity greater than 1) of the entity because indirect neighbors and each benchmark method for ranking social entities and
of an entity could be persuaded by the entity through entity predicting top persuaders for the second 10 weeks of the study
similarity or structural equivalence. Second, as discussed in period. The other part contained data over the second 10

weeks of the study period, which we used to calculate the tified for the second 10-week period. As shown, the top-K
persuasion credit of each social entity during this period. We precision of our method is substantially higher than that of
then assigned ranks to social entities according to their per- any benchmark method across K. For example, averaged
suasion credits and thereby identified top persuaders for the across K, the top-K precision of our method is 52.69% higher
second 10-week period. The performance of each method than that of eigenvector centrality, the best performing bench-
was evaluated by comparing top persuaders and ranks pre- mark method in terms of average top-K precision.6 It is also
dicted by the method against top persuaders and ranks worth noting that the top-K precision of random selection is
identified with the second 10 weeks of data, using the the worst among all the investigated methods, far below that
following metrics. of our method or any other benchmark method. In addition to
demonstrating the magnitude of our method’s improvement
We employed top-K precision (Manning et al. 2008) to com- over benchmark methods, we also applied the Wilcoxon
pare predicted top persuaders and top persuaders identified signed-ranks test to the performance data in Table 6 (Sned-
with the second 10 weeks of data. In our evaluation, the ecor and Cochran 1989) and noted that our method signi-
metric is calculated as ficantly outperformed each benchmark method in terms of
top-K precision (p <0.001).
|predicted top-K persuaders 1 identified top-K persuaders|
top-K precision =
K
We also compared our method and benchmark methods in
terms of the Spearman coefficient. As shown in Table 7, the
We also evaluated the extent to which ranks predicted by each
average Spearman coefficient of our method is 0.77, which
method agree or disagree with ranks computed with the
indicates a high, positive correlation between ranks predicted
second 10 weeks of data. A common measure is Spearman’s
by our method and ranks computed with the second 10 weeks
rank correlation coefficient that ranges from –1 (i.e., complete
of data. On the other hand, the average Spearman coefficient
disagreement) to 0 (i.e., independence) and then +1 (i.e.,
of intercentrality, the best performing benchmark method
complete agreement) (Kendall and Gibbons 1990; Snedecor
according to average Spearman coefficient, is 0.21, far below
and Cochran 1989). The higher the coefficient, the better
that of our method. The average Spearman coefficient of
ranks predicted by a method correlate with ranks computed
random selection is 0, which reveals that ranks predicted by
with the second 10 weeks of data (Kendall and Gibbons 1990;
this method are generally not correlated with ranks computed
Snedecor and Cochran 1989). Finally, it is interesting to
with the second 10 weeks of data at all. We further applied
assess to what extent each method achieves the objective of
the Wilcoxon signed-ranks test to the performance data in
the top persuader prediction problem (i.e., persuading as many
Table 7; the results showed that our method significantly
social entities to adopt as possible). We therefore computed
outperformed each benchmark method in terms of Spearman
total persuasion credit for a method as
coefficient (p < 0.001).
total persuasion credit =  pc(v)

v ∈VK
Table 8 summarizes total persuasion credits of our method
and those of each benchmark method across K. Taking K =
50, for example, the total persuasion credit of our method is
where VK denotes the set of top-K persuaders predicted by the 24.55, implying that the top 50 persuaders predicted by our
method and pc(v) is the persuasion credit of entity v. For each method take full credit for the virtual item purchase by
method, its total persuasion credit reflects the number of approximate 24 other avatars. Averaged across K, the total
social entities persuaded to adopt by the top persuaders pre- persuasion credit of our method is 111.49. On average, the
dicted by that method. total persuasion credit of our method is 56.37% higher than
that of eigenvector centrality, the best performing benchmark
method according to average total persuasion credit. This
Evaluation Results and Analyses improvement implies that top persuaders predicted by our
method could lead to a substantially larger number of avatars
Following the described procedure, we conducted our evalua- to purchase than those predicted by the best performing
tion, with γ = 0.5 and K varying from 50 (i.e., approximately benchmark method. Considering that the company from
1% of the total number of avatars) to 500 (i.e., approximately which we collected data gleans $45 million annually from
10% of the total number of avatars). In Table 6, we compared selling virtual items, such improvement could translate into
our method and benchmark methods in terms of top-K preci-
sion. The top-K precision of our method is 0.83, averaged 6
The improvement in top-K precision (e.g., 52.69%) is calculated as the top-K
across K; that is, on average, 83% of the top-K persuaders precision difference between our method and a benchmark method divided
predicted by our method are among top-K persuaders iden- by the top-K precision of the benchmark method.

Table 6. Performance Comparison on Top-K Precision (γ = 0.5)

Our Between- Intercentrality
K Method Degree Closeness ness Percolation Eigenvector (β = 0.1) INF-RANK INF-SIM Random
50 0.72 0.54 0.60 0.46 0.50 0.64 0.64 0.40 0.32 0.01
100 0.81 0.56 0.60 0.48 0.52 0.61 0.58 0.42 0.40 0.02
150 0.78 0.55 0.57 0.53 0.54 0.58 0.57 0.47 0.41 0.03
200 0.82 0.55 0.55 0.54 0.54 0.58 0.57 0.45 0.44 0.04
250 0.83 0.56 0.56 0.54 0.54 0.55 0.55 0.47 0.47 0.05
300 0.85 0.54 0.53 0.52 0.53 0.52 0.52 0.48 0.48 0.06
350 0.85 0.53 0.52 0.51 0.50 0.53 0.51 0.49 0.47 0.07
400 0.87 0.50 0.51 0.50 0.49 0.51 0.50 0.49 0.46 0.08
450 0.88 0.49 0.50 0.48 0.48 0.49 0.49 0.49 0.46 0.09
500 0.89 0.48 0.48 0.48 0.48 0.49 0.48 0.49 0.45 0.10
AVG 0.83 0.53 0.54 0.50 0.51 0.55 0.54 0.46 0.44 0.06
SD 0.05 0.03 0.04 0.03 0.02 0.05 0.05 0.03 0.05 0.03
Notes: We evaluated the performance of intercentrality with decay factor β ranging from 0.1 to 0.9, in increments of 0.1. Our method substantially
outperformed intercentrality in all three performance metrics, across all β values we investigated. For the interest of space, we only report the best
performing intercentrality in this paper. For example, in this table, intercentrality achieves the best average top-K precision when β = 0.1. For the
random selection method, the performances reported in Tables 6–8 as well as the tables in Appendixes B and D are the average performance of
running the method 1,000 times.
Table 7. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.5)

50 0.68 0.19 0.17 0.14 0.01 0.02 0.21 0.19 0.19 0.00
100 0.65 0.19 0.13 0.15 0.05 0.12 0.17 0.12 0.16 0.00
150 0.73 0.20 0.19 0.10 0.13 0.21 0.15 0.18 0.14 0.01
200 0.73 0.12 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.01
250 0.79 0.08 0.12 0.08 0.07 0.18 0.19 0.14 0.09 0.00
300 0.80 0.14 0.17 0.14 0.15 0.19 0.19 0.16 0.11 0.00
350 0.82 0.17 0.18 0.18 0.23 0.16 0.28 0.16 0.16 0.00
400 0.83 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.85 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.86 0.25 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.77 0.19 0.18 0.16 0.15 0.17 0.21 0.18 0.16 0.00
SD 0.07 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00

Table 8. Performance Comparison on Total Persuasion Credit (γ = 0.5)

16.91
50 24.55 16.81 17.34 17.40 18.40 19.18 18.80 15.68 2.89
100 47.00 31.33 32.58 30.46 32.29 33.41 32.31 31.97 28.59 5.73
150 67.80 43.49 44.78 44.45 44.94 45.08 45.73 44.61 39.84 8.48
200 87.78 56.95 55.30 57.23 57.67 58.00 57.93 54.89 51.84 11.51
250 106.42 69.11 68.03 68.70 69.31 66.84 65.89 66.90 63.26 14.23
300 124.16 78.10 76.73 77.73 78.85 76.11 75.60 77.43 73.72 17.02
350 140.97 86.87 86.01 85.88 85.44 86.89 84.25 87.88 81.69 19.88
400 157.30 92.17 94.03 92.61 92.45 93.93 92.35 95.38 88.41 22.91
450 172.30 100.67 101.37 99.25 101.02 101.14 101.11 102.78 95.65 25.63
500 186.62 108.30 107.70 106.51 108.92 109.22 108.00 110.10 101.80 28.42
AVG 111.49 68.38 68.39 68.02 68.93 68.98 68.20 68.89 64.05 15.67
SD 54.45 30.59 30.35 30.04 30.09 29.98 29.64 31.33 29.35 8.61
significant financial gains for the company. The total persua- three forces for top persuader prediction by demonstrating
sion credit of random selection remains the worst among all significant outperformance of our method over methods that
the evaluated methods. By applying the Wilcoxon signed- take only two of the three forces into account in Appendix C.
ranks test to the performance data in Table 8, we noted that Second, our method properly integrates these forces to predict
our method significantly outperformed each benchmark top persuaders, through the introduction of persuasion
method in total persuasion credit (p <0.001). probability. The effect of proper integration is manifested by
the substantial performance improvement of our method over
To ensure the robustness of our evaluation results, we per- INF-SIM, which multiplies social influence and entity
formed additional evaluations with different γ, ranging from similarity without due justification.
0.1 to 0.9, in increments of 0.1, except for γ = 0.5. By varying
γ from 0.1 to 0.9, we examined our method and benchmark Finally, we empirically analyzed key properties of our pro-
methods in different attenuation scenarios, from the most posed method, including its sensitivity to the number of
attenuation (γ = 0.1) to the least attenuation (γ = 0.9). We bootstrap samples and its convergence. Figure 3 plots top-K
obtained evaluation results largely similar to those in Tables precision of our method against the number of bootstrap
6–8 and reported detailed evaluation results in Appendix B. samples created by the method, ranging from 0 to 100, for K
Across the values we investigated, our method outperformed = 100 and γ = 0.5. We note that zero bootstrap sample means
the best performing benchmark method by 39.15% to 60.11% no bagging, whereas greater than zero bootstrap sample
in top-K precision, 256.53% to 280.33% in Spearman coeffi- indicates bagging is employed by our method. As shown in
cient, and 36.99% to 62.78% in total persuasion credit. Figure 3, the top-K precision of our method without bagging
Together, the evaluation results in Tables 6–8 and the addi- is 0.77, which is lower than that of our method with bagging
tional results in Appendix B consistently suggest the fol- regardless of the number of bootstrap samples created but is
lowing. First, our method substantially outperforms each substantially higher than that of the best performing bench-
benchmark method in all three evaluation metrics (i.e., top-K mark method (i.e., 0.61 according to Table 6). Furthermore,
precision, Spearman coefficient, and total persuasion credit), the top-K precision of our method attains the maximum when
across a wide range of K and γ. Second, ranks predicted by the number of bootstrap samples is 50. We varied K and γ
random selection have no correlation with ranks computed and observed largely similar results. We also studied the
with the second 10 weeks of data. This finding highlights the sensitivity of our method to the number of bootstrap samples,
need to develop effective methods to rank social entities and in terms of Spearman coefficient and total persuasion credit,
predict top persuaders, instead of relying on random selection. and observed similar results. Overall, our empirical results
The superiority of our method over benchmark methods can suggest that bagging can improve the performance of our
be attributed to two factors. First, our method is anchored in method, and that creating 50 bootstrap samples for bagging
relevant social network theories and considers three distinct seems reasonable, consistent with extant literature (Breiman
forces central to social persuasion suggested by these theories, 1996). Moreover, according to our empirical results, our
whereas most benchmark methods focus on social influence method substantially outperforms the best performing bench-
exclusively. We affirm the importance of considering all mark method, regardless of bagging or not bagging.

Figure 3. Top-K Precision of Our Method Against the Number of Bootstrap Samples Created
(K = 100, γ = 0.5)
Figure 4. Convergence of Our Method (γ = 0.5)
To analyze the convergence of our method, we ran the algo- tion method by drawing the knowledge and techniques from
rithm for predicting top persuaders (Figure 2) and plotted μ several disciplines that include machine learning and socio-
after each iteration of updating C in Figure 4. As shown, the logy. We empirically examine the effectiveness of our
value of μ becomes trivial and the algorithm converges after method with real-world social network data and demonstrate
four iterations for γ = 0.5. We varied γ from 0.1 to 0.9 and its superior performance over several prevalent methods.
noted the convergence of the algorithm after three or four
iterations for all γ. The observed convergence of the algo- Our study contributes to extant literature in several ways.
rithm is consistent with Theorem 1. First, we propose a novel method for predicting top per-
suaders in a social network. Our method simultaneously
considers several important forces central to social persuasion
Discussion and Conclusion (i.e., social influence, entity similarity, and structural equiv-
alence), while most existing methods focus on social influ-
In the networked economy, organizations strive to learn and ence. Second, we design an innovative way to properly
experiment with innovative ways of utilizing social networks integrate these forces in our method. According to salient
for business purposes. Toward that end, predicting top per- social network theories, social influence, entity similarity, and
suaders in a social network is critical. This study formulates structural equivalence jointly determine social persuasion.
the top persuader prediction problem and develops a novel We therefore introduce persuasion probability that denotes the
method to predict top persuaders. Our method is premised in likelihood of persuasion conditioned on these three essential
eminent social network theories (Burt 1987; Festinger 1954; forces. Persuasion probability thus functions as a nexus for
Knoke 1990; Weiner 1986), which serve as kernel theories to integrating these forces in our method. We propose how to
inform the design of our method by revealing several key estimate persuasion probabilities, design a power iteration
forces central to social persuasion (Gregor and Hevner 2013). algorithm to predict top persuaders using the estimated per-
We consider these forces to develop a top persuader predic- suasion probabilities, and analyze the theoretical property of

the algorithm. Finally, we demonstrate the effectiveness of to evaluate our proposed method. In such experiments, we
our proposed method by conducting an empirical evaluation can observe the number of followers after the adoption of top
involving real-world social network data. Overall, our evalu- persuaders predicted by our method and measure social
ation results show that our proposed method substantially influence, entity similarity, and structural equivalence from a
outperforms benchmark methods from representative prior top persuader to a follower. As a result, we can directly
research and salient industry practices. observe the performance of our method and, at the same time,
discover the mechanism leading to the observed performance.
Our study has implications for business and society as well. By combining our current evaluation results with archival
By analyzing massive data sets accumulated in organizations, data and field experimental outcomes, we are able to produce
our proposed method can effectively predict top persuaders more comprehensive and compelling empirical evidence
and therefore offers great value for various social network- suggesting the effectiveness and practical value of our
centric applications. For example, a firm could use our method. Another area worthy of future investigation is to
method to predict top persuaders among its potential cus- explore other algorithms for estimating persuasion proba-
tomers, entice them to adopt its product or service, and then bilities, such as Naïve Bayes or support vector machines.
leverage their high persuasion power to foster such adoptions Understandably, the accuracy of estimated persuasion
among many other customers. As shown in our evaluation probabilities directly affects the performance of our method.
results, top persuaders predicted by our method could lead to By experimenting with other probability estimation algo-
a much larger number of avatars purchasing a virtual item rithms, we can explore whether these algorithms could
than those predicted by any benchmark method. Given the enhance the accuracy of estimated persuasion probabilities, in
$14.8 billion annual global virtual goods market, which is comparison to the decision tree-based algorithm currently
expected to continue growing (Wohn 2014), the successful used in our method, and subsequently improve the perfor-
application of our method could generate significant and posi- mance of our method. Furthermore, susceptibility to persua-
tive financial impacts. The application of our method is not sion also seems to play an important role in the spread of
restricted to online games; rather, this method could be used adoptions (Aral and Walker 2012), in addition to persuasion.
for various product or service marketing in other industries. Therefore, it could be interesting to predict social entities that
are most susceptible to persuasion. Finally, the avatar data
Another potential application of our method is the promotion used in our evaluation have several limitations that deserve
of healthy lifestyles or practices, such as promoting a healthy future research attention. First, our evaluation results are
diet in a population, or fostering the acceptance of modern obtained using data spanning over the first 20-week period of
medicine in a population of an underdeveloped area. In these an online game. It would be interesting to further evaluate the
situations, we could construct a social network for the popu- performance of the proposed and benchmark methods with
lation in which a healthy lifestyle or practice is promoted and avatar data collected from other time periods or data sets from
use the proposed method to predict top persuaders in the a different application. Toward that end, we conducted an
social network. We would then target and convince these top additional evaluation with another data set and observed
persuaders to adopt the promoted healthy lifestyle or practice evaluation results suggesting the superiority of our method
and exploit their high persuasion power to promote it to the over the benchmark methods. We report detailed evaluation
whole population. In addition, the rapid growth of volum- results in Appendix D. Second, we did not control for the
inous social network data encourages innovative applications impacts of marketing efforts in our evaluation. Such efforts
of our method. An interesting case in point is to adapt our could impact the adoption behaviors of Avatars and communi-
method for infectious disease containment. In this case, cations among them, which in turn might affect our evaluation
individuals, if infected, who would cause wide-spread con- results. Future research should find appropriate ways to
tagion in a population, resemble the top persuaders we control for marketing efforts. Third, we use the profile of an
examined. Timely prediction and immunization of such con- avatar as a proxy for the profile of a real life user who creates
tagious individuals is indispensable to effective containment the avatar and implicitly assume that similarly profiled real
efforts, which requires proper adaptation of our proposed life users will create similarly profiled avatars. Future work
method to the infectious disease containment scenario. should collect profile data of real users to better measure
entity similarity.
There are several directions to extend our study. In our
evaluation, we follow a common practice to approximate true
top persuaders with top persuaders identified according to Acknowledgments
persuasion credit. Future work is needed to identify true top
persuaders for evaluating different top persuader prediction The authors thank the senior editor, the associate editor, and the
methods. It is also worthwhile to conduct field experiments three anonymous reviewers for their guidance and constructive

comments that have tremendously improved the paper. Xiao Fang Breiman, L. 1996. “Bagging Predictors,” Machine Learning (24:2)
is the corresponding author of the paper. pp. 123-140.
Brin, S., and Page, L. 1998. “The Anatomy of a Large-Scale
Hypertextual Web Search Engines,” Computer Networks and
References ISDN Systems (30:1-7), pp. 107-117.
Burt, R. S. 1987. “Social Contagion and Innovation: Cohesion
Agarwal, N., Liu, H., Tang, L., and Yu, P. S. 2008. “Identifying Versus Structural Equivalence,” American Journal of Sociology
the Influential Bloggers in a Community,” in Proceedings of the (92:6), pp. 1287-1335.
1st ACM International Conference on Web Search and Data Burt, R. S., Gabbay, S. M., Holt, G., and Moran, P. 1994.
Mining, Vancouver, October 26-31, pp. 207-217. “Contingent Organization as a Network Theory: The Culture-
Agarwal, R., Animesh, A., and Prasad, K. 2009. “Research Note- Performance Contingency Function,” Acta Sociologica (37:4),
Social Interactions and the ‘Digital Divide’: Explaining Varia- pp. 345-370.
tions in Internet Use,” Information Systems Research (20:2), pp. Carr, D. 2008. “How Obama Tapped into Social Networks’
277-294. Power,” The New York Times, November 9.
Albert, R., Jeong, H., and Barabasi, A. 2000. “Error and Attack Centola, D. 2010. “The Spread of Behavior in an Online Social
Tolerance of Complex Networks,” Nature (406), pp. 378-382. Network Environment,” Science (329:5996), pp. 1194-1197.
Anagnostopoulos, A., Kumar, R., and Mahdian, M. 2008. Cha, M., Haddadi, H., Benevenuto, F., and Gummadi, K. P. 2010.
“Influence and Correlation in Social Networks,” in Proceedings “Measuring User Influence in Twitter: The Million Follower
of the 14th ACM SIGKDD International Conference on Knowl- Fallacy,” in Proceedings of the 4th International AAAI Con-
edge Discovery and Data Mining, Las Vegas, NV, August 24-27, ference on Weblogs and Social Media, Washington, DC, May 23-
pp. 7-15. 26, pp. 10-17.
Aral, S., Muchnik, L., and Sundararajan, A. 2009. “Distinguishing Chaiken, S. L., Gruenfeld, D. H., and Judd, C. M. 2000. “Persu-
Influence Based Contagion from Homophily Driven Diffusion in asion in Negotiations and Conflict Situations,” The Handbook of
Dynamic Networks,” Proceedings of the National Academy of Conflict Resolution: Theory and Practice, M. Deutsch and P. T.
Sciences (106:51), pp. 21544-21549. Coleman (Eds.), San Francisco: Jossey-Bass, pp. 144-165.
Aral, S., and Walker, D. 2012. “Identifying Influential and Suscep- Crandall, D., Cosley, D., Huttenlocher, D., Kleinberg, J., and
tible Members of Social Networks,” Science (337:6092), pp. Suri, S. 2008. “Feedback Effects between Similarity and Social
337-341. Influence in Online Communities,” in Proceedings of the 14th
Bainbridge, W. S. 2007. “The Scientific Research Potential of ACM SIGKDD International Conference on Knowledge
Virtual Worlds,” Science (317:5837), pp. 472-476. Discovery and Data Mining, Las Vegas, NV, August 24-27, pp.
Bakshy, E., Hofman, J. M., Mason, W. A., and Watts, D. J. 2011. 160-168.
“Everyone’s an Influencer: Quantifying Influence on Twitter,” in Cussens, J. 1993. “Bayes and Pseudo-Bayes Estimates of Condi-
Proceedings of the 4th ACM International Conference on Web tional Probabilities and Their Reliability, ” in Proceedings of the
Search and Data Mining, Hong Kong, February 9-12, pp. 65-74. European Conference on Machine Learning, San Mateo, CA:
Ballester, C., Calvó-Armengol, A., and Zenou, Y. 2006. “Who’s Morgan Kaufmann, pp. 136-152.
Who in Networks. Wanted: The Key Player,” Econometrica Fang, X., Hu, P. J., Li, Z., and Tsai, W. 2013. “Predicting Adop-
(74:5), pp. 1403-1417. tion Probabilities in Social Networks,” Information Systems
Ballester, C., Calvó-Armengol, A., and Zenou, Y. 2010. Research (24:1), pp.128-145.
“Delinquent Networks,” Journal of the European Economic Festinger, L. 1954. “A Theory of Social Comparison Processes,”
Association (8:1), pp. 34-61. Human Relations (7:2), pp.117-140.
Bass, F. 1969. “A New Product Growth for Model Consumer Forbes Insights. 2011. “Bringing 20/20 Foresight to Marketing,”
Durables,” Management Science (15:5), pp. 215-227. May (http://www.forbes.com/forbesinsights/bringing_foresight_
Bavelas, A. 1948. “A Mathematical Model for Group Structures,” to_marketing/).
Human Organization (7:3), pp. 16-30. Freeman, L. C. 1979. “Centrality in Social Networks: Conceptual
Bonacich, P. 1972. “Factoring and Weighing Approaches to Status Clarification,” Social Networks (1:3), pp. 215-239.
Score and Clique Identification,” Journal of Mathematical Friedkin, N. E. 1998. A Structural Theory of Social Influence,
Sociology (2:1), pp. 113-120. New York: Cambridge University Press.
Bonacich, P. 1987. “Power and Centrality: A Family of Mea- Ghosh, R., and Lerman, K. 2010. “Predicting Influential Users in
sures,” American Journal of Sociology (92:5), pp. 1170-1182. Online Social Networks,” in Proceedings of the 4th KDD Work-
Bonacich, P. 2007. “Some Unique Properties of Eigenvector shop on Social Network Analysis, Washington, DC, July 25-28.
Centrality,” Social Networks (29:4), pp. 555-564. Golub, G. H., and Van Loan, C. F. 1996. Matrix Computations (3rd
Borgatti, S. P., Mehra, A., Brass, D. J., and Labianca, G. 2009. ed.), Baltimore, MD: Johns Hopkins University Press.
“Network Analysis in the Social Sciences,” Science (323:5916), Granovetter, M. 1978. “Threshold Models of Collective Behavior,”
pp. 892-895. American Journal of Sociology (83:6), pp. 1420-1443.
Brass, D. J. 1984. “Being in the Right Place: A Structural Analysis Gregor, S., and Hevner, A. 2013. “Positioning and Presenting
of Individual Influence in an Organization,” Administrative Design Science Research for Maximum Impact,” MIS Quarterly
Science Quarterly (29:4), pp. 518-539. (37:2), pp. 337-355.

Heath, M. T. 2002. Scientific Computing: An Introductory Survey Social Information Processing,” Administrative Science
(2nd ed.), New York: McGraw-Hill. Quarterly (36:2), pp. 219-244.
Hemp, P. 2006. “Avatar-Based Marketing,” Harvard Business Rice, R. E., Grand, A. E., Schmitz, J., and Torobin, J. 1990.
Review (84:6), pp. 48-57. “Individual and Network Influences on the Adoption and Per-
Jackson, M. O., and Wolinsky, A. 1996. “A Strategic Model of ceived Outcomes of Electronic Messaging,” Social Networks
Social and Economic Networks,” Journal of Economic Theory (12:1), pp. 27-55.
(71:1), pp. 44-74. Romero, D. M., Galuba, W., Asur, S., and Huberman, B. A. 2010.
Katz, L. 1953. “A New Status Index Derived from Sociometric “Influence and Passivity in Social Media,” Working Paper, HP
Analysis,” Psychometrika (18:1), pp. 39-43. Labs.
Kempe, D., Kleinberg, J., and Tardos, E. 2003. “Maximizing the Shalizi C. R., and Thomas, A. C. 2011. “Homophily and Contagion
Spread of Influence through a Social Network,” in Proceedings are Generically Confounded in Observational Social Network
of the 9th ACM SIGKDD International Conference on Knowledge Studies,” Social Methods Research (40:2), pp. 211-239.
Discovery and Data Mining, Washington, DC, August 24-27, Simon, H. A. 1976. Administrative Behavior, New York: Free
pp.137-146. Press.
Kendall, M. G., and Gibbons, J. D. 1990. Rank Correlation Snedecor, G. W., and Cochran, W. G. 1989. Statistical Methods,
Methods, Oxford, UK: Oxford University Press. Ames, IA: Blackwell Publishing Professional.
Kleinberg, J. 2007. “Cascading Behavior in Networks: Algorith- Sulzmann, J. N., and Furnkranz, J. 2009. “An Empirical Com-
mic and Economic Issues,” in Algorithmic Game Theory, parison of Probability Estimation Techniques for Probabilistic
Cambridge, UK: Cambridge University Press. Rules,” in Proceedings of the 12th International Conference on
Kleinberg, J. 2008. “The Convergence of Social and Technological Discovery Science, Porto, Portugal, October 3-5, pp. 317-331.
Networks,” Communications of the ACM (51:11), pp. 66-72. Susarla, A., Oh, J. H., and Tan, Y. 2012. “Social Networks and the
Knoke, D. 1990. “Networks of Political Action: Toward Theory Diffusion of User-generated Content: Evidence from YouTube,”
Construction,” Social Forces (68:4), pp. 1041-1063. Information Systems Research (23:1), pp. 23-41.
Lorrain, F., and White, H. C. 1971. “Structural Equivalence of Tan, P. N., Steinbach, M., and Kumar, V. 2006. Introduction to
Individuals in Social Networks,” Journal of Mathematical Data Mining, Boston: Pearson-Addison Wesley.
Sociology (1:1), pp. 49-80. Travers, J., and Milgram, S. 1969. “An Experimental Study of the
Manning, C. D., Raghavan, P., and Schütze, H. 2008. Introduction Small World Problem,” Sociometry (32:4), pp. 425-443.
to Information Retrieval, Cambridge, UK: Cambridge University Trusov, M., Bodapati, A. V., and Bucklin, R. E. 2010. “Deter-
Press. mining Influential Users in Internet Social Networks,” Journal of
Mitchell, T. M. 1997. Machine Learning, New York: McGraw- Marketing Research (47:4), pp. 643-658.
Hill. (New chapters can be accessed at http://www.cs.cmu.edu/ Valente, T. W. 1995. Network Models of the Diffusion of Innova-
~tom/NewChapters.html.) tions, Cresskill, NJ: Hampton Press.
Pazzani, M., Merz, C., Murphy, P., Ali, K., Hume, T., and Brunk, C. Walter, W. P., Koput, K. W., and Smith-Doerr, L. 1996. “Inter-
1994. “Reducing Misclassification Costs, ” in Proceedings of the organizational Collaboration and the Locus of Innovation:
11th International Conference of Machine Learning, Boca Raton, Networks of Learning in Biotechnology,” Administrative Science
FL, December 12-15, pp. 217-225. Quarterly (41:1), pp. 116-145.
Petty, R. E., and Cacioppo, J. T. 1986. Communication and Wasserman, S., and Faust, K. 1994. Social Network Analysis:
Persuasion: Central and Peripheral Routes to Attitude Change, Methods and Applications, Cambridge, UK: Cambridge
New York: Springer-Verlag. University Press.
Pfeffer, J., Salancik, G. R., and Leblebici, H. 1976. “The Effect of Watts, A. 2001. “A Dynamic Model of Network Formation,”
Uncertainty on the Use of Social Influence in Organizational Games and Economic Behavior (34:1), pp. 331-341.
Decision Making,” Administrative Science Quarterly (21:2), pp. Watts, D. J., and Dodds, P. S. 2007. “Influentials, Networks, and
227-245. Public Opinion Formation,” Journal of Consumer Research
Piraveenan, M., Prokopenko, M., and Hossain, L. 2013. “Perco- (34:4), pp. 441-458.
lation Centrality: Quantifying Graph-Theoretic Impact of Nodes Watts, D. J., and Peretti, J. 2007. “Viral Marketing for the Real
During Percolation in Networks,” PLoS One (8:1), pp. 1-14. World,” Harvard Business Review (85:5), pp. 22-23.
Probst, F., Grosswiele, L., and Pfleger, R. 2013. “Who Will Lead Weiner, B. 1974. Achievement Motivation and Attribution Theory,
and Who Will Follow: Identifying Influential Users in Online Morristown, NJ: General Learning Press.
Social Networks—A Critical Review and Future Research Weiner, B. 1986. An Attributional Theory of Motivation and
Directions,” Business and Information Systems Engineering Emotion, New York: Springer Verlag.
(5:3), pp. 179-193. Wejnert, B. 2002. “Integrating Models of Di usion of Innovations:
Quinlan, J. R. 1993. C4.5: Programs for Machine Learning, San A Conceptual Framework,” Annual Review of Sociology (28), pp.
Francisco: Morgan Kaufmann Publishers. 297-326.
Rice, R. E., and Aydin, C. 1991. “Attitudes toward New Organi- Weng, J., Lim, E., Jiang, J., and He, Q. 2010. “TwitterRank:
zational Technology: Network Proximity as a Mechanism for Finding Topic-sensitive Influential Twitterers,” in Proceedings

of the 3rd ACM International Conference on Web Search and About the Authors
Data Mining, New York, February 3-6, pp. 261-270.
White, D. R., and Borgatti, S. P. 1994. “Betweenness Centrality Xiao Fang is an associate professor of Management Information
Measures for Directed Graphs,” Social Networks (16:4), pp. Systems at the Lerner College of Business and Economics, Univer-
335-346. sity of Delaware. He received his B.S. and M.S. degrees from
Wohn, D. 2014. “Spending Real Money: Purchasing Patterns of Fudan University, China, and his Ph.D. in Management Information
Virtual Goods in an Online Social Game,” in Proceedings of the Systems from the University of Arizona. He studies business and
2014 ACM SIGCHI Conference on Human Factors in Computing social network analytics with research methods and tools drawn
Systems, Toronto, April 26-May 1, pp. 3359-3368. from reference disciplines including Management Science (e.g.,
Wood, W. 2000. “Attitude Change: Persuasion and Social Optimization) and Computer Science (e.g., Data Mining and
Influence,” Annual Review of Psychology (51), pp. 539-570. Machine Learning). He has published in MIS Quarterly, Manage-
Zadrozny, B., and Elkan, C. 2001. “Learning and Making Decisions ment Science, Information Systems Research, Operations Research,
When Costs and Probabilities Are Both Unknown,” in INFORMS Journal on Computing, Journal of Management Infor-
Proceedings of the 7th International Conference on Knowledge mation Systems, ACM Transactions on Information Systems, IEEE
Discovery and Data Mining, San Francisco, August 26-29, pp. Transactions on Knowledge and Data Engineering, among others.
204-213.
Paul Jen-Hwa Hu is David Eccles Chair Professor at the David
Zanden, J. W. 1987. Social Psychology (4th ed.), New York:
Eccles School of Business, the University of Utah. He received his
Random House.
Ph.D. in Nanagement Information Systems from the University of
Zhao, K., Yen, J., Greer, G., Qiu, B., Mitra, P., and Portier, K.
Arizona. His current research interests include information tech-
2014. “Finding Influential Users of Online Health Communities:
nology for health care, technology implementation management,
A New Metric Based on Sentiment Influence,” Journal of the
digital transformation of business and government, technology
American Medical Informatics Association (21:e2), pp. e212-218.
impacts and evaluations, and knowledge management. Paul has
Zhuge, H., and Zhang, J. 2010. “Topological Centrality and its published in MIS Quarterly, Information Systems Research, Journal
e-Science Applications,” Journal of the American Society for of Management Information Systems, Journal of the AIS, Decision
Informatics Science and Technology (61:9), pp. 1824-1841. Sciences, and various IEEE and ACM transactions.

RESEARCH ARTICLE
TOP PERSUADER PREDICTION FOR SOCIAL NETWORKS

Xiao Fang
Lerner College of Business and Economics, University of Delaware,
Newark, DE 19716-3130 U.S.A. {[email protected]}
Paul Jen-Hwa Hu
David Eccles School of Business, University of Utah,
Salt Lake City, UT 84112 U.S.A. {[email protected]}
Appendix A
Proof of Theorem 1
Our proof is built on Perron-Frobenius theorem, a seminal work in matrix theory (Meyer 2000). By Perron-Frobenius theorem, the power
iteration algorithm for predicting top K persuaders converges to a unique C and this convergence is independent of the initialization of C if
the persuasion probability matrix P is nonnegative, irreducible, and aperiodic (Heath 2002). We first show that P is nonnegative. Each
component of the right hand side of Equation (10) is positive except nD $ 0; thus, persuasion probability pij estimated with Equation (10) is
positive, for all i, j = 1, 2, …, n and i … j. Because all diagonal elements of P are equal to zero and all non-diagonal elements of P are positive
persuasion probabilities, P is nonnegative.
We now prove that P is irreducible and aperiodic. Let G(P) be the directed graph associated with P. According to Meyer (2000), G(P) is
defined as a directed graph with n nodes {N1, N2, …, Nn}, and there exists an edge from Ni to Nj if element pij of P is positive, where i, j = 1,
2, …, n and i … j. To prove P is irreducible and aperiodic is equivalent to show that G(P) is strongly connected and aperiodic (Meyer 2000).
We first show G(P) is strongly connected. A directed graph is strongly connected if for every ordered pair Ni, Nj of its nodes there exists a path
from Ni to Nj (West 2001). Since pij > 0 for all i, j = 1, 2, …, n and i … j, by the definition of G(P), there is an edge from Ni to Nj for any ordered
pair Ni, Nj of G(P). That is, for any ordered pair Ni, Nj of G(P), there is a path of length 1 from Ni to Nj. Therefore, G(P) is strongly connected.
The period of a directed graph is the greatest common divisor of the lengths of its cycles and a directed graph is aperiodic if its period is 1
(Denardo 1977; Jarvis and Shier 1999). As discussed above, there exists an edge from Ni to Nj for any ordered pair Ni, Nj of G(P). We thus
list cycles that start and end at N1 and their respective lengths in Table A1.
Apparently, the greatest common divisor of the cycle lengths in Table A1 is 1. Table A1 only lists part of the cycles in G(P). Hence, the
greatest common divisor of the lengths of all the cycles in G(P) must be 1, and by definition, G(P) is aperiodic. This completes the proof.
MIS Quarterly Vol. 42 No. 2—Appendices/March 2018 A1

Table A1. Some Cycles in G(P) and Their Lengths
References
Denardo, E. V. 1977. “Periods of Connected Networks and Powers of Nonnegative Matrices,” Mathematics of Operations Research (2:1),
pp. 20-24.
Heath, M. T. 2002. Scientific Computing: An Introductory Survey (2nd ed.), New York: McGraw-Hill.
Jarvis, J. P., and Shier, D. R. 1999. “Graph-Theoretic Analysis of Finite Markov Chains,” in Applied Mathematical Modeling: A
Multidisciplinary Approach, Boca Raton, FL: CRC Press.
Meyer, C. D. 2000. Matrix Analysis and Applied Linear Algebra, Philadelphia, PA: Society for Industrial and Applied Mathematics.
West, D. B. 2001. Introduction to Graph Theory (2nd ed.), Upper Saddle River, NJ: Prentice Hall.
A2 MIS Quarterly Vol. 42 No. 2—Appendices/March 2018

Appendix B
Additional Evaluation Results
We conducted additional evaluations with γ ranging from 0.1 to 0.9 in increments of 0.1, except for γ = 0.5. As summarized in Tables B1 to
B24, our method substantially outperformed each benchmark method in all three evaluation metrics, across γ and K.
Table B1. Performance Comparison on Top-K Precision (γ = 0.1)

Our Between- Intercentrality INF-
K Method Degree Closeness ness Percolation Eigenvector (β = 0.1) RANK INF-SIM Random
50 0.70 0.56 0.58 0.50 0.52 0.60 0.58 0.42 0.36 0.01
100 0.78 0.57 0.58 0.49 0.53 0.58 0.55 0.45 0.41 0.02
150 0.77 0.56 0.57 0.53 0.55 0.59 0.58 0.49 0.42 0.03
200 0.76 0.57 0.56 0.55 0.55 0.57 0.56 0.47 0.46 0.04
250 0.75 0.57 0.55 0.55 0.54 0.53 0.52 0.48 0.49 0.05
300 0.75 0.55 0.55 0.54 0.54 0.54 0.54 0.51 0.49 0.06
350 0.78 0.55 0.54 0.52 0.51 0.55 0.53 0.52 0.48 0.07
400 0.80 0.52 0.54 0.53 0.50 0.53 0.53 0.51 0.48 0.08
450 0.79 0.52 0.52 0.52 0.51 0.52 0.52 0.51 0.48 0.09
500 0.80 0.51 0.51 0.51 0.50 0.52 0.51 0.50 0.48 0.10
AVG 0.77 0.55 0.55 0.52 0.52 0.55 0.54 0.49 0.45 0.06
SD 0.03 0.02 0.02 0.02 0.02 0.03 0.02 0.03 0.04 0.03
Table B2. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.1)

50 0.54 0.18 0.15 0.12 0.00 0.01 0.21 0.16 0.18 -0.01
100 0.60 0.19 0.13 0.15 0.04 0.12 0.16 0.12 0.16 0.00
150 0.67 0.18 0.19 0.10 0.14 0.20 0.14 0.18 0.14 0.00
200 0.71 0.12 0.20 0.09 0.11 0.14 0.15 0.19 0.11 0.00
250 0.76 0.08 0.12 0.08 0.08 0.17 0.19 0.14 0.09 0.00
300 0.78 0.14 0.17 0.14 0.15 0.19 0.18 0.16 0.11 0.00
350 0.77 0.17 0.17 0.18 0.23 0.16 0.28 0.16 0.16 0.00
400 0.78 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.82 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.84 0.26 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.73 0.18 0.18 0.16 0.15 0.16 0.21 0.18 0.16 0.00
SD 0.10 0.06 0.04 0.07 0.09 0.07 0.05 0.04 0.05 0.00

Table B3. Performance Comparison on Total Persuasion Credit (γ = 0.1)

50 43.92 31.24 32.05 31.93 33.31 35.05 34.37 29.39 27.45 2.87
100 77.81 54.96 56.55 52.03 55.39 57.62 55.54 51.63 47.11 5.69
150 104.84 74.15 75.30 74.01 73.75 75.99 75.84 70.52 62.78 8.60
200 127.70 92.10 89.93 91.35 91.29 93.11 92.11 83.48 79.64 11.49
250 147.22 107.75 106.42 105.84 105.21 105.02 102.98 97.88 95.19 14.13
300 163.48 119.04 117.42 116.70 117.39 116.85 115.97 111.76 108.96 17.22
350 179.63 130.61 128.23 126.67 124.86 129.27 125.82 125.39 117.92 19.96
400 192.71 136.62 138.02 134.54 133.14 137.40 135.15 134.18 126.94 22.76
450 202.45 146.50 146.56 142.56 142.62 145.85 145.52 142.63 135.41 25.83
500 210.62 155.08 153.40 150.29 151.34 154.46 153.05 150.10 141.22 28.45
AVG 145.04 104.81 104.39 102.59 102.83 105.06 103.64 99.70 94.26 15.70
SD 55.85 41.03 40.32 39.69 38.90 39.46 39.23 40.48 38.92 8.63

50 0.72 0.54 0.56 0.48 0.50 0.62 0.60 0.40 0.34 0.01
100 0.79 0.57 0.59 0.49 0.53 0.59 0.57 0.44 0.41 0.02
150 0.77 0.55 0.57 0.53 0.55 0.59 0.58 0.49 0.42 0.03
200 0.80 0.57 0.57 0.56 0.55 0.58 0.58 0.47 0.46 0.04
250 0.79 0.57 0.56 0.55 0.55 0.54 0.54 0.48 0.48 0.05
300 0.80 0.56 0.55 0.53 0.53 0.54 0.54 0.51 0.49 0.06
350 0.83 0.55 0.54 0.53 0.51 0.55 0.53 0.52 0.49 0.07
400 0.83 0.51 0.52 0.51 0.49 0.52 0.51 0.50 0.47 0.08
450 0.84 0.50 0.50 0.50 0.49 0.50 0.50 0.49 0.46 0.09
500 0.84 0.50 0.49 0.49 0.49 0.50 0.50 0.50 0.46 0.10
AVG 0.80 0.54 0.55 0.52 0.52 0.55 0.54 0.48 0.45 0.06
SD 0.04 0.03 0.03 0.03 0.03 0.04 0.04 0.04 0.05 0.03

50 0.59 0.19 0.16 0.13 0.00 0.01 0.21 0.18 0.18 0.00
100 0.65 0.19 0.13 0.15 0.04 0.12 0.16 0.12 0.16 0.00
150 0.70 0.17 0.19 0.10 0.14 0.20 0.14 0.18 0.14 0.00
200 0.72 0.12 0.20 0.09 0.10 0.14 0.16 0.19 0.11 0.00
250 0.78 0.08 0.12 0.08 0.08 0.17 0.19 0.14 0.09 0.00
300 0.78 0.14 0.17 0.14 0.15 0.19 0.19 0.16 0.11 0.00
350 0.80 0.17 0.17 0.18 0.23 0.16 0.28 0.16 0.16 0.00
400 0.81 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.84 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.86 0.26 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.75 0.18 0.18 0.16 0.15 0.16 0.21 0.18 0.16 0.00
SD 0.09 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 36.02 25.15 25.93 25.74 27.06 28.49 27.97 24.25 22.48 2.84
100 65.59 45.27 46.86 43.17 45.94 47.89 46.29 43.72 39.54 5.74
150 90.76 61.92 63.20 62.00 62.27 63.75 64.09 60.15 53.34 8.43
200 113.39 78.29 76.47 77.92 78.07 79.59 79.10 72.23 68.39 11.38
250 132.90 92.97 91.86 91.39 91.48 90.47 88.96 85.78 82.52 14.27
300 150.89 103.63 102.14 101.73 102.78 101.57 100.86 98.60 95.23 17.08
350 167.05 114.11 112.68 111.14 109.91 113.77 110.58 111.12 103.96 19.87
400 181.46 120.02 122.03 118.73 117.80 121.68 119.70 119.52 111.98 22.88
450 193.51 129.55 130.31 126.35 127.08 129.87 129.64 127.64 120.08 25.82
500 204.28 138.02 137.21 134.02 135.65 138.55 137.19 135.18 126.20 28.45
AVG 133.59 90.89 90.87 89.22 89.81 91.56 90.44 87.82 82.37 15.68
SD 56.46 37.52 37.13 36.39 35.91 36.48 36.16 37.25 35.45 8.65

50 0.72 0.56 0.58 0.48 0.52 0.64 0.62 0.42 0.34 0.01
100 0.84 0.58 0.61 0.50 0.53 0.61 0.58 0.44 0.42 0.02
150 0.79 0.55 0.57 0.53 0.54 0.58 0.57 0.47 0.42 0.03
200 0.83 0.56 0.56 0.55 0.55 0.58 0.58 0.46 0.46 0.04
250 0.81 0.56 0.56 0.55 0.54 0.54 0.54 0.48 0.48 0.05
300 0.83 0.55 0.54 0.52 0.53 0.54 0.53 0.49 0.48 0.06
350 0.84 0.54 0.53 0.52 0.51 0.54 0.52 0.51 0.48 0.07
400 0.85 0.51 0.52 0.51 0.50 0.52 0.51 0.50 0.47 0.08
450 0.86 0.50 0.50 0.49 0.49 0.50 0.50 0.49 0.46 0.09
500 0.86 0.49 0.48 0.48 0.48 0.49 0.49 0.49 0.45 0.10
AVG 0.82 0.54 0.55 0.51 0.52 0.55 0.54 0.47 0.45 0.06
SD 0.04 0.03 0.04 0.02 0.02 0.05 0.04 0.03 0.04 0.03
Table B8. Performance Comparison on Spearman’s Rank Corrrelation Coefficient (γ = 0.3)

50 0.57 0.19 0.16 0.13 0.01 0.01 0.21 0.18 0.18 0.00
100 0.64 0.19 0.13 0.15 0.04 0.12 0.16 0.12 0.16 0.00
150 0.75 0.17 0.19 0.10 0.13 0.21 0.14 0.18 0.14 0.00
200 0.75 0.12 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.78 0.08 0.12 0.08 0.07 0.17 0.19 0.14 0.09 0.00
300 0.81 0.14 0.17 0.14 0.15 0.19 0.19 0.16 0.11 0.00
350 0.85 0.17 0.17 0.18 0.23 0.16 0.28 0.16 0.16 0.00
400 0.86 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.88 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.89 0.26 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.78 0.18 0.18 0.16 0.15 0.16 0.21 0.18 0.16 0.00
SD 0.10 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 30.93 21.43 22.12 22.01 23.21 24.37 23.92 21.00 19.46 2.86
100 57.84 39.14 40.62 37.60 39.96 41.58 40.21 38.64 34.76 5.72
150 81.23 54.08 55.23 54.35 54.78 55.67 56.20 53.45 47.47 8.57
200 103.27 69.12 67.40 69.04 69.32 70.38 70.11 64.82 61.25 11.31
250 123.06 82.84 81.78 81.73 82.09 80.45 79.21 77.77 74.27 14.30
300 141.13 92.89 91.46 91.59 92.70 90.87 90.26 89.69 86.11 17.15
350 157.21 102.67 101.57 100.52 99.62 102.60 99.63 101.37 94.57 19.94
400 172.49 108.38 110.45 107.81 107.17 110.22 108.41 109.43 102.02 22.78
450 186.30 117.56 118.40 115.06 116.21 118.07 117.93 117.28 109.79 25.53
500 198.69 125.76 125.15 122.62 124.57 126.60 125.29 124.78 116.00 28.36
AVG 125.22 81.39 81.42 80.23 80.96 82.08 81.12 79.82 74.57 15.65
SD 56.27 34.75 34.47 33.85 33.59 33.95 33.61 34.87 32.96 8.59
Table 10. Performance Comparison on Top-K Precision (γ = 0.4)

50 0.74 0.54 0.60 0.46 0.50 0.64 0.64 0.40 0.32 0.01
100 0.82 0.57 0.60 0.49 0.52 0.61 0.58 0.43 0.41 0.02
150 0.78 0.55 0.57 0.52 0.53 0.58 0.57 0.47 0.41 0.03
200 0.82 0.56 0.56 0.55 0.55 0.58 0.58 0.46 0.45 0.04
250 0.81 0.56 0.56 0.54 0.54 0.55 0.54 0.48 0.47 0.05
300 0.85 0.55 0.54 0.52 0.53 0.53 0.53 0.49 0.48 0.06
350 0.85 0.54 0.53 0.52 0.51 0.54 0.52 0.50 0.48 0.07
400 0.87 0.50 0.51 0.51 0.49 0.51 0.50 0.49 0.46 0.08
450 0.87 0.50 0.50 0.49 0.48 0.50 0.50 0.50 0.46 0.09
500 0.87 0.49 0.48 0.48 0.48 0.49 0.49 0.49 0.45 0.10
AVG 0.83 0.53 0.54 0.51 0.51 0.55 0.54 0.47 0.44 0.06
SD 0.04 0.03 0.04 0.03 0.02 0.05 0.05 0.03 0.05 0.03

50 0.61 0.19 0.17 0.14 0.01 0.02 0.21 0.19 0.19 0.00
100 0.63 0.19 0.13 0.15 0.04 0.12 0.17 0.12 0.16 -0.01
150 0.73 0.18 0.19 0.10 0.13 0.21 0.14 0.18 0.14 0.00
200 0.72 0.12 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.79 0.08 0.12 0.08 0.07 0.17 0.19 0.14 0.09 -0.01
300 0.80 0.14 0.17 0.14 0.15 0.19 0.19 0.16 0.11 0.00
350 0.83 0.17 0.17 0.18 0.23 0.16 0.28 0.16 0.16 0.00
400 0.83 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.86 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.87 0.25 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.77 0.18 0.18 0.16 0.15 0.17 0.21 0.18 0.16 0.00
SD 0.09 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 27.34 18.80 19.41 19.39 20.48 21.43 21.02 18.70 17.32 2.89
100 51.67 34.73 36.09 33.57 35.64 36.98 35.77 34.90 31.29 5.73
150 73.52 48.36 49.37 48.79 49.27 49.73 50.34 48.50 43.19 8.53
200 94.58 62.32 60.64 62.45 62.81 63.47 63.33 59.29 55.99 11.47
250 113.56 75.20 74.14 74.47 74.98 72.88 71.81 71.74 68.14 14.32
300 132.04 84.69 83.29 83.90 85.02 82.68 82.14 82.90 79.23 17.13
350 148.50 93.92 92.98 92.42 91.77 93.93 91.14 93.91 87.44 19.94
400 164.64 99.42 101.41 99.41 99.02 101.26 99.57 101.68 94.48 22.97
450 178.92 108.26 109.04 106.35 107.83 108.78 108.69 109.30 101.97 25.62
500 192.41 116.17 115.58 113.76 115.95 117.09 115.83 116.71 108.17 28.59
AVG 117.72 74.19 74.19 73.45 74.28 74.82 73.96 73.76 68.72 15.72
SD 55.49 32.48 32.24 31.79 31.69 31.82 31.47 32.95 30.99 8.64
Table B13. Performance Comparison on Top-K Precision (γ - 0.6)

50 0.72 0.56 0.60 0.46 0.52 0.66 0.64 0.40 0.32 0.01
100 0.80 0.56 0.60 0.48 0.52 0.61 0.58 0.42 0.40 0.02
150 0.78 0.55 0.57 0.53 0.54 0.58 0.57 0.47 0.41 0.03
200 0.84 0.55 0.56 0.54 0.54 0.58 0.58 0.44 0.44 0.04
250 0.83 0.55 0.54 0.54 0.53 0.54 0.54 0.46 0.46 0.05
300 0.86 0.54 0.53 0.52 0.53 0.52 0.52 0.48 0.48 0.06
350 0.86 0.52 0.51 0.50 0.50 0.52 0.50 0.49 0.46 0.07
400 0.88 0.49 0.50 0.49 0.48 0.50 0.49 0.48 0.46 0.08
450 0.89 0.48 0.48 0.48 0.47 0.48 0.48 0.48 0.45 0.09
500 0.89 0.48 0.47 0.48 0.48 0.48 0.47 0.48 0.45 0.10
AVG 0.83 0.53 0.54 0.50 0.51 0.55 0.54 0.46 0.43 0.06
SD 0.05 0.03 0.04 0.03 0.03 0.06 0.05 0.03 0.05 0.03
Table B14. Performance Comparison on Spearman’s Rank Correlation Coeffiient (γ = 0.6)

50 0.59 0.19 0.17 0.15 0.01 0.02 0.21 0.20 0.19 0.00
100 0.63 0.19 0.13 0.15 0.05 0.12 0.17 0.13 0.16 0.01
150 0.71 0.20 0.20 0.10 0.13 0.21 0.15 0.18 0.14 -0.01
200 0.73 0.12 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.78 0.08 0.12 0.08 0.08 0.18 0.19 0.14 0.09 0.00
300 0.79 0.14 0.17 0.14 0.15 0.20 0.19 0.16 0.11 0.00
350 0.81 0.17 0.18 0.18 0.23 0.17 0.28 0.16 0.16 0.00
400 0.83 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.85 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.86 0.25 0.25 0.26 0.24 0.23 0.28 0.24 0.22 0.00
AVG 0.76 0.19 0.19 0.16 0.15 0.17 0.22 0.18 0.16 0.00
SD 0.09 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 22.29 15.22 15.70 15.82 16.73 17.39 17.03 15.47 14.36 2.84
100 43.20 28.59 29.75 27.95 29.58 30.52 29.51 29.58 26.39 5.74
150 62.92 39.90 41.06 40.93 41.42 41.31 41.98 41.41 37.09 8.61
200 82.07 52.56 50.93 52.97 53.45 53.50 53.48 51.25 48.43 11.40
250 100.29 64.07 62.99 63.92 64.61 61.84 61.00 62.88 59.22 14.23
300 117.73 72.62 71.27 72.60 73.71 70.64 70.17 72.86 69.13 17.09
350 134.63 80.98 80.18 80.42 80.16 80.99 78.49 82.81 76.88 20.05
400 150.94 86.10 87.84 86.90 86.95 87.77 86.27 90.08 83.34 22.72
450 166.56 94.30 94.89 93.27 95.31 94.70 94.69 97.28 90.34 25.61
500 181.40 101.65 101.01 100.38 102.99 102.54 101.36 104.51 96.43 28.66
AVG 106.20 63.60 63.56 63.52 64.49 64.12 63.40 64.81 60.16 15.70
SD 53.48 28.92 28.69 28.52 28.71 28.38 28.04 29.94 27.95 8.64
Table B16. Performance Comparison on Top-K Precision (γ - 0.7)

50 0.70 0.54 0.60 0.44 0.50 0.64 0.64 0.38 0.30 0.01
100 0.78 0.56 0.61 0.48 0.53 0.61 0.59 0.42 0.40 0.02
150 0.78 0.55 0.57 0.53 0.54 0.58 0.57 0.47 0.41 0.03
200 0.84 0.55 0.56 0.54 0.54 0.58 0.58 0.44 0.44 0.04
250 0.84 0.55 0.54 0.54 0.54 0.54 0.54 0.46 0.46 0.05
300 0.86 0.53 0.52 0.51 0.52 0.51 0.51 0.47 0.47 0.06
350 0.88 0.51 0.51 0.49 0.49 0.51 0.49 0.48 0.45 0.07
400 0.88 0.48 0.49 0.49 0.47 0.49 0.48 0.48 0.45 0.08
450 0.89 0.48 0.48 0.47 0.47 0.48 0.48 0.48 0.45 0.09
500 0.90 0.47 0.46 0.47 0.47 0.47 0.46 0.47 0.44 0.10
AVG 0.84 0.52 0.53 0.50 0.51 0.54 0.53 0.46 0.43 0.06
SD 0.06 0.03 0.05 0.03 0.03 0.06 0.06 0.03 0.05 0.03

50 0.60 0.19 0.17 0.15 0.01 0.02 0.21 0.20 0.19 -0.01
100 0.67 0.20 0.13 0.15 0.05 0.13 0.17 0.13 0.16 0.00
150 0.69 0.20 0.20 0.10 0.13 0.21 0.15 0.18 0.14 0.00
200 0.71 0.12 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.77 0.08 0.13 0.08 0.08 0.18 0.19 0.14 0.09 0.00
300 0.81 0.14 0.17 0.14 0.15 0.20 0.19 0.16 0.11 0.00
350 0.80 0.17 0.18 0.18 0.23 0.17 0.28 0.16 0.16 0.00
400 0.82 0.25 0.20 0.24 0.26 0.19 0.26 0.20 0.21 0.00
450 0.83 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.85 0.25 0.25 0.26 0.24 0.23 0.28 0.23 0.22 0.00
AVG 0.76 0.19 0.19 0.16 0.15 0.17 0.22 0.18 0.16 0.00
SD 0.08 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00

Table B18. Performance Companion on Total Persuasion Credit (γ = 0.7)

50 20.47 13.92 14.34 14.52 15.36 15.91 15.58 14.27 13.26 2.91
100 39.93 26.32 27.40 25.85 27.33 28.12 27.18 27.57 24.55 5.73
150 58.80 36.90 37.94 37.98 38.46 38.15 38.83 38.71 34.77 8.63
200 77.19 48.86 47.25 49.37 49.88 49.70 49.72 48.16 45.53 11.43
250 94.98 59.81 58.72 59.86 60.61 57.61 56.85 59.44 55.78 14.23
300 112.05 67.95 66.63 68.22 69.33 65.99 65.55 68.93 65.20 17.01
350 129.04 75.95 75.18 75.75 75.64 75.94 73.55 78.45 72.74 19.94
400 145.35 80.89 82.51 81.99 82.24 82.47 81.04 85.50 78.98 22.84
450 161.39 88.80 89.31 88.12 90.41 89.13 89.15 92.52 85.78 25.61
500 176.74 95.90 95.22 95.08 97.87 96.75 95.60 99.65 91.79 28.43
AVG 101.59 59.53 59.45 59.67 60.71 59.98 59.30 61.32 56.84 15.67
SD 52.55 27.46 27.23 27.19 65.25 26.96 63.68 28.71 26.73 8.59

Our Intercentrality INF-
K Method Degree Closeness Betweenness Percolation Eigenvector (β = 0.1) RANK INF-SIM Random
50 0.72 0.54 0.60 0.44 0.50 0.64 0.64 0.40 0.30 0.01
100 0.79 0.56 0.60 0.48 0.52 0.61 0.58 0.43 0.40 0.02
150 0.78 0.54 0.57 0.53 0.53 0.58 0.57 0.47 0.40 0.03
200 0.84 0.55 0.56 0.54 0.54 0.58 0.58 0.44 0.44 0.04
250 0.84 0.55 0.54 0.54 0.53 0.54 0.54 0.46 0.46 0.05
300 0.87 0.53 0.52 0.51 0.52 0.51 0.51 0.47 0.47 0.06
350 0.88 0.51 0.50 0.49 0.49 0.51 0.49 0.47 0.45 0.07
400 0.90 0.48 0.49 0.48 0.47 0.48 0.47 0.47 0.45 0.08
450 0.91 0.46 0.47 0.46 0.46 0.47 0.46 0.47 0.44 0.09
500 0.92 0.46 0.45 0.46 0.46 0.46 0.45 0.47 0.43 0.10
AVG 0.84 0.52 0.53 0.49 0.50 0.54 0.53 0.45 0.42 0.06
SD 0.06 0.04 0.05 0.04 0.03 0.06 0.06 0.02 0.05 0.03
Table B20. Performance Comparison on Spearman’s Rank Corellation Coefficient (γ = 0.8)

50 0.60 0.19 0.17 0.15 0.01 0.02 0.21 0.20 0.19 0.00
100 0.65 0.20 0.13 0.15 0.06 0.13 0.17 0.13 0.16 0.00
150 0.69 0.21 0.20 0.10 0.13 0.21 0.15 0.18 0.14 0.00
200 0.72 0.13 0.20 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.77 0.09 0.13 0.08 0.07 0.18 0.19 0.14 0.09 0.00
300 0.80 0.15 0.17 0.14 0.15 0.20 0.19 0.16 0.11 0.00
350 0.82 0.17 0.18 0.18 0.23 0.17 0.28 0.15 0.16 0.00
400 0.83 0.25 0.20 0.24 0.26 0.19 0.27 0.20 0.21 0.00
450 0.85 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.21 0.00
500 0.86 0.25 0.25 0.26 0.24 0.23 0.28 0.23 0.22 0.00
AVG 0.76 0.19 0.19 0.17 0.15 0.17 0.22 0.18 0.16 0.00
SD 0.09 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 18.92 12.83 13.21 13.42 14.20 14.66 14.35 13.25 12.33 2.90
100 37.29 24.39 25.40 24.07 25.41 26.07 25.20 25.84 22.98 5.73
150 55.24 34.35 35.28 35.46 35.92 35.46 36.13 36.38 32.76 8.57
200 72.92 45.68 44.09 46.26 46.81 46.44 46.47 45.47 43.02 11.46
250 90.26 56.12 55.02 56.34 57.15 53.96 53.26 56.43 52.79 14.27
300 107.16 63.89 62.59 64.40 65.53 61.95 61.54 65.48 61.77 17.16
350 123.93 71.57 70.82 71.66 71.71 71.53 69.25 74.63 69.12 19.88
400 140.40 76.34 77.85 77.69 78.13 77.84 76.47 81.47 75.17 22.78
450 156.53 83.99 84.41 83.59 86.11 84.25 84.29 88.32 81.76 25.77
500 172.34 90.84 90.13 90.40 93.37 91.64 90.53 95.35 87.70 28.50
AVG 97.50 56.00 55.88 56.33 57.43 56.38 55.75 58.26 53.94 15.70
SD 51.60 26.15 25.93 25.99 26.42 25.69 25.37 27.61 25.64 8.62

50 0.72 0.54 0.60 0.44 0.50 0.64 0.64 0.40 0.30 0.01
100 0.79 0.56 0.61 0.48 0.53 0.61 0.59 0.43 0.40 0.02
150 0.78 0.53 0.57 0.52 0.53 0.58 0.58 0.47 0.40 0.03
200 0.83 0.55 0.56 0.54 0.54 0.58 0.58 0.44 0.44 0.04
250 0.84 0.55 0.54 0.54 0.53 0.54 0.54 0.46 0.46 0.05
300 0.87 0.53 0.52 0.51 0.52 0.51 0.51 0.47 0.47 0.06
350 0.88 0.51 0.50 0.49 0.49 0.51 0.49 0.47 0.45 0.07
400 0.89 0.48 0.49 0.48 0.47 0.48 0.47 0.47 0.44 0.08
450 0.91 0.46 0.47 0.45 0.46 0.47 0.46 0.46 0.44 0.09
500 0.92 0.45 0.45 0.45 0.46 0.46 0.45 0.46 0.43 0.10
AVG 0.84 0.52 0.53 0.49 0.50 0.54 0.53 0.45 0.42 0.06
SD 0.06 0.04 0.06 0.04 0.03 0.06 0.06 0.02 0.05 0.03

50 0.59 0.19 0.17 0.15 0.01 0.02 0.22 0.21 0.19 0.01
100 0.63 0.20 0.14 0.15 0.06 0.13 0.17 0.14 0.16 0.00
150 0.68 0.21 0.21 0.10 0.13 0.22 0.15 0.18 0.15 0.00
200 0.72 0.13 0.21 0.09 0.10 0.14 0.16 0.20 0.11 0.00
250 0.76 0.09 0.13 0.08 0.07 0.18 0.19 0.14 0.09 0.00
300 0.80 0.15 0.17 0.14 0.15 0.20 0.19 0.16 0.11 0.00
350 0.81 0.18 0.18 0.18 0.23 0.17 0.28 0.15 0.16 0.00
400 0.82 0.25 0.20 0.24 0.26 0.19 0.27 0.19 0.21 0.00
450 0.84 0.25 0.23 0.26 0.24 0.23 0.26 0.23 0.20 0.00
500 0.86 0.26 0.25 0.26 0.24 0.23 0.28 0.23 0.22 0.00
AVG 0.75 0.19 0.19 0.17 0.15 0.17 0.22 0.18 0.16 0.00
SD 0.09 0.06 0.04 0.07 0.09 0.06 0.05 0.04 0.05 0.00


50 17.59 11.89 12.24 12.48 13.20 13.59 13.30 12.37 11.53 2.85
100 34.93 22.73 23.68 22.52 23.75 24.31 23.48 24.33 21.60 5.73
150 52.10 32.13 32.97 33.27 33.71 33.12 33.79 34.33 31.00 8.51
200 69.11 42.91 41.33 43.55 44.12 43.58 43.64 43.10 40.81 11.40
250 86.00 52.87 51.78 53.24 54.10 50.76 50.10 53.76 50.15 14.33
300 102.67 60.32 59.04 61.03 62.17 58.40 58.01 62.42 58.74 16.94
350 119.29 67.69 66.97 68.04 68.23 67.62 65.45 71.22 65.90 19.99
400 135.75 72.32 73.72 73.87 74.48 73.72 72.41 77.87 71.77 22.73
450 152.06 79.71 80.05 79.56 82.30 79.90 79.96 84.56 78.18 25.60
500 168.22 86.33 85.59 86.22 89.37 87.08 86.00 91.49 84.04 28.56
AVG 93.77 52.89 52.74 53.38 54.54 53.21 52.61 55.55 51.37 15.67
SD 50.67 24.97 24.75 24.91 25.44 24.54 24.23 26.61 24.65 8.62
Appendix C
Analysis of Three Versus Two Forces for Top Persuader Prediction
Our method integrates three important forces (i.e., social influence, entity similarity, and structural equivalence) to predict top persuaders. To
demonstrate the importance and value of considering three forces for top persuader prediction, we conducted additional analyses using the data
and evaluation design detailed in the “Empirical Evaluation and Results” section. Specifically, we removed structural equivalence from our
method and persuasion probability originally defined in Equation (8) became
(
pij = P D j = 1| Iij , γ
dij − 1
Mij ) (C1)
We labeled this new method without structural equivalence as R-removed. Similarly, we removed entity similarity only and labeled the method
without entity similarity as M-removed; we dropped social influence only and labeled the method without social influence as I-removed. The
performance differences between our method and these new methods, each of which considers only two forces, reveal the need to consider three
forces for top persuader prediction; they also shed light on the contribution of each removed force to the performance of our method.
In Tables C1–C3, we report the performance comparisons between our method and each method that only considers two forces, with γ = 0.5.
Our method substantially outperforms R-removed, M-removed, and I-removed in each evaluation metric, across the K values we investigated.
Averaged across K, our method outperforms R-removed, M-removed, and I-removed by 20.28%, 17.66%, and 68.85%, respectively, in terms
of top-K precision and by 76.60%, 99.68%, and 483.31%, respectively, in terms of Spearman coefficient. On average, our method is 22.72%
higher in total persuasion credit than R-removed, 20.21% higher than M-removed and 32.34% higher than I-removed. We further analyzed
different values ranging from 0.1 to 0.9, in increments of 0.1, except for γ = 0.5, and observed results largely similar to those reported in Tables
C1–C3. Across , our method outperforms R-removed by 11.95% to 27.20% in top-K precision, 71.14% to 258.14% in Spearman coefficient,
and 13.60% to 27.25% in total persuasion credit. Similarly, our method outperforms M-removed by 11.80% to 19.88% in top-K precision,
66.38% to 204.92% in Spearman coefficient, and 13.41% to 23.54% in total persuasion credit; it also outperforms I-removed by 66.22% to
80.35% in top-K precision, 128.09% to 537.94% in Spearman coefficient, and 31.82% to 40.31% in total persuasion credit. Overall, our method
significantly outperforms the methods that consider only two forces, suggesting the necessity of considering three forces for top persuader
prediction. Our results also indicate that each of the three forces contributes to the performance of our method; in particular, social influence
seems to contribute the most among the three forces.

Table C1. Performance Comparison on Top=K Precision (γ = 0.5)

K Our Method R-removed M-removed I-removed
50 0.72 0.60 0.64 0.30
100 0.81 0.69 0.69 0.43
150 0.78 0.71 0.72 0.50
200 0.82 0.71 0.73 0.52
250 0.83 0.72 0.73 0.47
300 0.85 0.71 0.73 0.55
350 0.85 0.70 0.72 0.56
400 0.87 0.70 0.71 0.58
450 0.88 0.68 0.69 0.58
500 0.89 0.68 0.69 0.55
AVG 0.83 0.69 0.71 0.50
SD 0.05 0.03 0.03 0.09
Table C2. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.5)

50 0.68 0.26 0.19 0.02
100 0.65 0.24 0.25 0.19
150 0.73 0.38 0.37 0.20
200 0.73 0.41 0.41 0.17
250 0.79 0.45 0.46 0.37
300 0.80 0.50 0.43 0.29
350 0.82 0.56 0.48 0.35
400 0.83 0.63 0.53 0.39
450 0.85 0.68 0.53 0.38
500 0.86 0.70 0.55 0.42
AVG 0.77 0.48 0.42 0.28
SD 0.07 0.16 0.12 0.13
Table C3. Performance Comparison on Total Persuasion Credit (γ = 0.5)

50 24.55 21.61 23.56 21.24
100 47.00 41.44 41.48 38.50
150 67.80 58.51 59.81 54.96
200 87.78 74.02 76.10 70.19
250 106.42 87.99 88.60 78.13
300 124.16 99.70 101.28 94.01
350 140.97 109.73 112.48 104.24
400 157.30 122.08 122.90 112.39
450 172.30 131.72 132.96 120.10
500 186.62 141.30 143.28 124.08
AVG 111.49 88.81 90.25 81.78
SD 54.45 39.66 39.82 35.36

Appendix D
Empirical Evaluation with Another Data Set
We conducted an additional evaluation with another data set. The data set contains 6.01 million records of phone communications among
28,440 mobile phone users over 20 weeks. Each record corresponds to a phone communication and consists of the timestamp and duration
of the communication as well as the respective identities of the two users participating in the communication. We can construct a social network
from these data. A social entity of the network represents a mobile phone user, a relationship between two entities exists if there are phone
communications between their corresponding users, and the strength of social interactions between two entities is measured as the
communication time between their corresponding users. The data set also contains adoption information: whether a user adopted a particular
mobile phone service during the study period (i.e., initial purchase of the service) and, if adopted, in which week. Over the study period, a total
of 3,129 users adopted the service. We also have data about each user’s profile, including gender, age, and membership levels in the two most
recent years.
Following the same procedure described in the “Evaluation Design” subsection, we used the data over the first 10 weeks of the study period
to train our method and each benchmark method for predicting top-K persuaders, where K varies from 280 (i.e., approximately 1% of the total
number of users) to 2,800 (i.e., approximately 10% of the total number of users), in increments of 280. We then employed data over the second
10 weeks of the study period to evaluate the prediction performance of each method. In Tables D1–D3, we report the performance of our
method and each benchmark method in terms of top-K precision, Spearman’s rank correlation coefficient, and total persuasion credit
respectively, with attenuation factor γ = 0.5. As shown, our method substantially outperforms all the benchmark methods in each performance
metric, across the investigated K values. Across K, our method, on average, is 312.01% higher in top-K precision than eigenvector centrality
(the best performing benchmark method in terms of average top-K precision), 82.04% higher in the Spearman coefficient than eigenvector
centrality (the best performing benchmark method according to average Spearman coefficient), and 207.68% higher in total persuasion credit
than intercentrality (the best performing benchmark method in terms of average total persuasion credit). In addition, we applied the Wilcoxon
signed-ranks test to the performance data in these tables and noted that our method significantly outperformed each benchmark method in any
performance metric (p < 0.001). To ensure the robustness of our evaluation results, we conducted more evaluations with different γ, ranging
from 0.1 to 0.9, in increments of 0.1, except for γ = 0.5. We obtained evaluation results largely similar to those in Tables D1–D3. Overall,
across the values we investigated, our method outperforms the best performing benchmark method by a range of 208.59% to 379.07% in top-K
precision, 57.48% to 110.65% in Spearman coefficient, and 115.01% to 263.26% in total persuasion credit.
Table D1. Performance Comparison on Top-K Precision (γ = 0.5)

280 0.80 0.16 0.05 0.14 0.16 0.21 0.17 0.09 0.06 0.01
560 0.80 0.17 0.08 0.14 0.14 0.19 0.18 0.12 0.09 0.02
840 0.78 0.15 0.10 0.14 0.14 0.16 0.17 0.12 0.09 0.03
1120 0.76 0.15 0.10 0.14 0.14 0.15 0.16 0.13 0.11 0.04
1400 0.73 0.15 0.11 0.13 0.14 0.15 0.15 0.12 0.11 0.05
1680 0.70 0.16 0.12 0.14 0.15 0.15 0.16 0.13 0.12 0.06
1960 0.66 0.17 0.13 0.14 0.16 0.17 0.17 0.13 0.12 0.07
2240 0.63 0.16 0.13 0.14 0.16 0.17 0.17 0.14 0.13 0.08
2520 0.59 0.16 0.14 0.15 0.16 0.17 0.18 0.15 0.13 0.09
2800 0.58 0.18 0.16 0.16 0.17 0.20 0.20 0.16 0.14 0.10
AVG 0.70 0.16 0.11 0.14 0.15 0.17 0.17 0.13 0.11 0.05
SD 0.08 0.01 0.03 0.01 0.01 0.02 0.01 0.02 0.03 0.03

Table D2. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.5)

280 0.52 0.10 0.04 0.15 0.01 0.22 0.36 0.13 0.12 0.00
560 0.52 0.23 -0.06 0.04 0.09 0.48 0.40 0.03 0.08 0.00
840 0.55 0.22 -0.03 0.06 0.07 0.49 0.35 0.08 0.01 0.00
1120 0.55 0.19 0.04 0.06 0.12 0.43 0.32 0.07 -0.03 0.00
1400 0.58 0.16 -0.02 0.09 0.09 0.37 0.31 0.10 0.00 0.00
1680 0.59 0.18 0.11 0.07 0.11 0.35 0.28 0.08 0.03 0.00
1960 0.61 0.19 0.01 0.10 0.11 0.31 0.27 0.09 0.05 0.00
2240 0.63 0.21 -0.02 0.09 0.09 0.30 0.24 0.07 0.03 0.00
2520 0.63 0.20 0.00 0.08 0.10 0.28 0.22 0.05 0.04 0.00
2800 0.63 0.20 -0.01 0.09 0.14 0.24 0.22 0.07 0.04 0.00
AVG 0.58 0.19 0.01 0.08 0.09 0.35 0.30 0.08 0.04 0.00
SD 0.04 0.04 0.05 0.03 0.04 0.10 0.06 0.03 0.04 0.00
Table D3. Performance Comparison on Total Persuasion Credit (γ = 0.5)

280 195.17 58.00 20.30 37.24 43.30 84.94 67.65 35.38 25.29 11.56
560 341.35 94.54 43.79 71.46 74.63 115.48 110.64 66.81 46.14 22.99
840 458.04 121.89 69.19 100.37 101.51 134.01 140.84 88.65 68.43 34.27
1120 550.64 150.69 92.49 124.65 128.15 157.92 165.08 111.75 95.96 46.11
1400 617.37 183.13 112.53 141.36 155.97 180.09 185.52 129.74 116.02 56.80
1680 677.77 202.90 140.09 165.47 181.58 199.38 207.60 150.31 131.69 67.07
1960 715.98 220.63 166.56 185.38 206.50 223.65 230.23 168.08 147.83 79.90
2240 750.42 234.20 184.71 205.66 227.57 241.07 250.54 191.25 171.27 91.15
2520 780.36 250.02 207.77 229.83 252.31 258.07 276.12 215.75 188.05 101.87
2800 804.90 269.20 232.11 253.92 262.54 282.53 299.61 229.00 204.89 113.87
AVG 589.20 178.52 126.95 151.53 163.41 187.71 193.38 138.67 119.56 62.56
SD 202.53 70.23 71.25 69.76 75.62 64.83 73.96 64.12 60.60 34.30

Misq 42 1 63

Uploaded by

Copyright:

Available Formats

Misq 42 1 63

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Misq 42 1 63

Uploaded by

Copyright:

Available Formats

RESEARCH ARTICLE

TOP PERSUADER PREDICTION FOR SOCIAL NETWORKS1

Introduction1 social networks is social persuasion that refers to the prin-

DOI: 10.25300/MISQ/2018/13211 MIS Quarterly Vol. 42 No. 1, pp. 63-82/March 2018 63

64 MIS Quarterly Vol. 42 No. 1/March 2018

MIS Quarterly Vol. 42 No. 1/March 2018 65

Table 1. Comparison of Our Method and Existing Methods

66 MIS Quarterly Vol. 42 No. 1/March 2018

network; the cohesion model focuses on socialization between Problem Formulation

MIS Quarterly Vol. 42 No. 1/March 2018 67

Equation (1) measures social influence from vi to vj as the ( )

68 MIS Quarterly Vol. 42 No. 1/March 2018

probability as the probability that vj adopts, given the impacts

MIS Quarterly Vol. 42 No. 1/March 2018 69

We therefore propose to estimate pij using bagged decision tree

70 MIS Quarterly Vol. 42 No. 1/March 2018

(a) Training (b) Probability Estimation

Figure 1. Bagged Decision Tree Induction with M-Estimation

Figure 2. Predicting Top K Persuaders

MIS Quarterly Vol. 42 No. 1/March 2018 71

Table 3. Summary of Avatar Profile Attributes

72 MIS Quarterly Vol. 42 No. 1/March 2018

MIS Quarterly Vol. 42 No. 1/March 2018 73

Table 4. Summary of Benchmark Methods

Table 5. An Illustration of Influence Score and Persuasion Credit

74 MIS Quarterly Vol. 42 No. 1/March 2018

total persuasion credit =  pc(v)

MIS Quarterly Vol. 42 No. 1/March 2018 75

Table 6. Performance Comparison on Top-K Precision (γ = 0.5)

Table 7. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.5)

76 MIS Quarterly Vol. 42 No. 1/March 2018

Table 8. Performance Comparison on Total Persuasion Credit (γ = 0.5)

MIS Quarterly Vol. 42 No. 1/March 2018 77

Figure 4. Convergence of Our Method (γ = 0.5)

78 MIS Quarterly Vol. 42 No. 1/March 2018

MIS Quarterly Vol. 42 No. 1/March 2018 79

80 MIS Quarterly Vol. 42 No. 1/March 2018

MIS Quarterly Vol. 42 No. 1/March 2018 81

82 MIS Quarterly Vol. 42 No. 1/March 2018

TOP PERSUADER PREDICTION FOR SOCIAL NETWORKS

MIS Quarterly Vol. 42 No. 2—Appendices/March 2018 A1

Table A1. Some Cycles in G(P) and Their Lengths

A2 MIS Quarterly Vol. 42 No. 2—Appendices/March 2018

Table B1. Performance Comparison on Top-K Precision (γ = 0.1)

Table B2. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.1)

MIS Quarterly Vol. 42 No. 2—Appendices/March 2018 A3

Table B3. Performance Comparison on Total Persuasion Credit (γ = 0.1)

Table B4. Performance Comparison on Top-K Precision (γ = 0.2)

Table B5. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.2)

A4 MIS Quarterly Vol. 42 No. 2—Appendices/March 2018

Table B6. Performance Comparison on Total Persuasion Credit (γ = 0.2)

Table B7. Performance Comparison on Top-K Precision (γ = 0.3)

Table B8. Performance Comparison on Spearman’s Rank Corrrelation Coefficient (γ = 0.3)

MIS Quarterly Vol. 42 No. 2—Appendices/March 2018 A5

Table B9. Performance Comparison on Total Persuasion Credit (γ = 0.3)

Table 10. Performance Comparison on Top-K Precision (γ = 0.4)

Table B11. Performance Comparison on Spearman’s Rank Correlation Coefficient (γ = 0.4)

A6 MIS Quarterly Vol. 42 No. 2—Appendices/March 2018

Table B12. Performance Comparison on Total Persuasion Credit (γ = 0.4)