A Diffusion Mechanism for Online Advertising Service
Over Social Media
Yung-Ming Li
and Ya-Lin Shiu
Institute of Information Management, National ChiaoTung University, Hsinchu, Taiwan
Abstract. Social media has increasingly become a popular platform for diffus-
ing information, through the message sharing of numerous participants in a so-
cial network. Recently, companies attempt to utilize social media to expose
their advertisements to appropriate customers. The success of message propaga-
tion in social media highly depends on the content relevance and closeness of
social relationships. In this paper, considering the factors of user preference,
network influence, and propagation capability, we propose a social diffusion
mechanism to discover the appropriate and influential endorsers from the social
network to deliver relevant advertisements broadly. The proposed mechanism
is implemented and verified in one of the most famous micro-blogging system-
Plurk. Our experimental results shows that the proposed model could efficiently
enhance the advertising exposure coverage and effectiveness.
1 Introduction
In recent years, social media has been flourished and raised high much popularity and
attention. Social media provides a great platform to diffuse information through the
numerous populations. Common social media marketing tools include Twitter, You-
tube, Facebook and so on. An overwhelming majority (88%) of marketers are using
social media to market their businesses, and a significant 81% of marketers indicate
that their efforts in social media have generated effective exposure for their business-
es, according to a social media study by Michael Stelzner[12].In 2010, one of the
most popular micro-blogging websites, Twitter, announced an innovative advertising
model, ” Promoted Tweets”. Promoted Tweets makes tweets as ads, which are dis-
tinctive from both traditional search ads and recent social ads. They measure the
advertising performance and the payment of sponsored tweets by “resonance” - the
interactions between users and a particular sponsored tweet such as retweet, reply, or
bookmarking [13]. How to choose the right people to deliver the information, how to
take the advantage of the social media, and how to design an ads diffusion mechan-
ism to widen the spreading coverage are crucial issues in the online advertising cam-
paign. In this paper, to address these issues, we design a social media diffusion me-
chanism, based on the concepts of content recommendation and network routing.
Once the appropriate messages and diffusion paths are identified, message can be
effectively delivered with support of the generated information. Considering the fac-
tors of user preference, network influence, and propagation capability, our system can
Li Y. and Shiu Y. (2010).
A Diffusion Mechanism for Online Advertising Service Over Social Media.
In Proceedings of the 4th International Workshop on Architectures, Concepts and Technologies for Service Oriented Computing, pages 42-51
DOI: 10.5220/0003049500420051
Copyright
c
SciTePress
effectively identify the most appropriate nodes in the social network for delivering
the relevant ads and recommend the friends for information sharing for an interme-
diate node.
2 Related Literature
2.1 Social Media
Social media are Internet platforms designed to disseminate information or messages
through social interaction, using highly accessible and scalable publishing techniques.
Social media is composed of content (information) and social interaction interface
(intimate community engagement and social viral activity). With its emerging trend
and promising popularity, researchers have put academic efforts in analyzing the
characteristics and functionalities of social media. For example, Kaplan and Haenlein
[6] examine the challenges and opportunities of social media and recommend a set of
ten rules that companies should follow when developing their own social media strat-
egy. To effectively communicate with customers, researchers engaged in analyzing
marketing trends and social relations. Gilbert and Karahalios [4] develop a predictive
model that maps data of social activity to tie strength so as to improve design ele-
ments of social media.To better figure out the users’ behaviors, many researchers
analyze the social influence, social interactions, and information diffusion in social
media[3].Comparing to the existing works, the study of information diffusion me-
chanism design of social media is apparently rare and new.
2.2 Online Advertising
The issue of online advertising has aroused much academic interests and been spot-
lighted for decades. Online advertising usually could be categorized into two types: 1)
targeting advertising, which deliver the ads based on user’s preference profiles, 2)
social advertising, which deliver the deliver the ads the influential users determined
by social relationship[11]. Targeted advertising usually applies the content-based and
collaborative-based approaches to discover users’ personal preferences. Compared
with the traditional online advertising, social advertising is a form of advertisement
that addresses people as part of a social network and uses the social relations and
social influences between people for selling products [14]. Some researchers measure
the influential strength by analyzing the number of network links and users’ relation
and interaction in the network to identify the influential nodes for social advertis-
ing[11, 14]. In this paper, considering the factors of user preference, network influ-
ence, and propagation capability, we propose a social diffusion mechanism to identify
the appropriate and influential endorsers from the social network to deliver relevant
advertisements broadly.
17
2.3 Information Diffusion and Social Routing
Researchers analyze information diffusion in the social network based on individual’s
characteristics. Some based on the bond percolation, graph theory or probabilistic
model to extract the influential nodes, considering the aspect of dynamic characteris-
tics, such as distance, time, and interaction and so on [7-8]. By revealing influential
factors and realizing the processes of the information diffusion, marketers can predict
when and how the information spreads over social networks to maximize the ex-
pected spreading result [5]. In this paper, we include static and dynamic factors di-
mensions to evaluate the propagation ability of nodes in social network.The design of
social diffusion mechanism is conceptually similar to that of computer network
routing process in selecting paths to switching the packet. Routing directs packets to
be forwarded from their source toward their ultimate destination through intermediate
nodes; hardware devices usually called routers. However, in the context of social
network, the links in social networks are formed by social relations and interaction
and researchers focus on the study of the issue: delivering the right information to the
right nodes and spreading widely. The goal can be achieved by implementing feasible
approaches to discover the influential nodes and leveraging the social relations to
diffuse the information between users further.In our paper, we incorporate the con-
cept of network routing to develop a social endorser engine to generate “social
routing tables” to support the information diffusion in a social network.
3 System Architecture
Analogous to the routing process in computer networks, we design a social diffusion
mechanism which sends a recommended list of users to our initial nodes with suitable
path for information diffusion. The recommendation lists suggest the users who have
strong propagation abilities in social networks. The users are referred as social en-
dorsers potentially willing to transmit the information to all his/her friends.
Notice that the proposed social diffusion mechanism is different from spamming.
We recommended those friends based on their preferences, social influence, and
propagation abilities via quantitative measurement. The advertising message will be
guided to right people by user’s judgments with the support of the system recommen-
dation. If users deliver ads to their friends, it means that users also think their friends
like the ads. The mechanism takes the advantage of content relevance and social
relation to reduce the negative impression of the advertisement and gain the advertis-
ing effectiveness. Social media provided us the source data of individual’s preference,
social relation and social influence. Preference is an important issue in target advertis-
ing. The social influence between users of the social media happened when they are
affected by others. It is likely that we are usually influenced by our friends or our
family. The social relation is a crucial factor to empower the social influence. If a
user frequently interacts with someone, to some extent, there are more similarities
and closeness between them. Therefore, we incorporate these components into our
proposed social endorser discovery engine.
18
3.1 Social Endorser Discovery Engine
Effective information diffusion on social networks is grounded on the relevance of
individual preference and closeness of social relations. Therefore, the main functio-
nality of the proposed social endorser discovery engine is to identify the nodes with
the strong propagation ability in disseminating relevant messages as wider as possi-
ble. In order to identify the appropriate social endorsers so as to achieve a better dif-
fusion performance, in this research, we not only consider the static factor, but also
dynamic factor in the evaluation of nodes’ diffusion capability - transmit information
towards the most suitable friends and spread widely further. Figure 1 displays the
main components and procedures of our social endorser discovery engine.
Fig. 1. Architecture of endorser discovery engine.
User Preference Analysis Module
As people trend to share information interesting to the receiver, discovering users’
preference is an important factor to be considered in the social endorser discovery. By
analyzing users’ preference, we can better understand what kinds of information are
suitable to be shared between the users. In order to realize user’s affinity levels of
information categories, we adopt a tree-like structure to categorize a set of informa-
tion and users’ preference. Tree-like structure is practically employed in many re-
searches, such as product taxonomy [1, 9].Besides, we utilize a distance-based ap-
proach to calculate the similarity between user categories and information categories.
The preference of a user and the type of an advertising message are described by a
catalog node they belong. Assume and stand for the category 1 (a user’s prefe-
rence) and catalog 2 (an advertisement type) belong to respectively and represent the
catalog of the first mutual parent nodes of catalogs 1 and 2. The fitness degree of the
ads to a user can be calculated by the following formula:
()
2
,
12
2
12
D
fn
Sim C C
P
DD D
fn
=
++
(1)
19
Network Influence Analysis Module
Connection Degree Influence. For the purpose of evaluating the relative importance
of user position in the whole network, social network analysis is applied. Degree
centrality is defined as the number of direct connections/links upon a node. Specifi-
cally, in-degree is a count of the number of connections directed to the node, and out-
degree is the number of connections that the node directs to others. In this research,
first, we consider the spammer or bots attempt to follow many people in order to gain
attention. Secondly, based on in-degree or out-degree ignores the ability for a user to
interact with content in the social network. Therefore, we use mutual relation (friend-
ship) to measure the degree centrality as in practice, mutual degree represents the
number of friends a user has. Mutual degree for user i is measured as
()
1
ij
n
MD i
j
E
=
=
(2)
where is 1 if an edge from node i to j exists and an edge from node j to node i ex-
ists, too, otherwise it is 0.
Content Degree Influence. Content Degree Influence is used to evaluate the popu-
larity of what a user posts. We measure the content degree of a user by the count of
the total responses and message forwards by people. We denoted as the total number
of the elements in a set. The formula for content degree measure can be expressed as:
() ()
()
||| |
()
||
reponse i forwrad i
CD i
post i
Φ+Φ
=
Φ
(3)
where
()
post i
Φ
stands for a set of the messages posted by user i,
()
reponse i
Φ
represents the set of the responses on user i ‘s posts , and
()
f
orward i
Φ
is the
set of i’s posts forwarded by others. The aggregate network influence score is the
sum of the mutual degree value
(
)
M
Di
and the content degree value
()CD i
.
Propagation Strength Analysis Module
Social Similarity. Social similarity aims to measure the similarity of two people from
implicit social structure and behaviors, such as “friend-in-common” and “content-in-
common”. The more friends-in-common of two people generally reflects the higher
connection level between them. If two people have more common friends, their inter-
ests should be more similar and the possibility that a people will forward a message
he feels interesting to the other becomes higher. Denote
(
)
i
as a set of user i ’s
friends. The similarity of friend-in-common between user i and friend j , is measured
as:
(
)
(
)
() ( )
()
||
(, )
||,| |
Fi F j
Sim i j
CF
M
ax F i F j
=
(4)
In addition, the more content-in-common posted by two people, the higher similar-
ity degree between them. Semantic analysis can be use to evaluate the social similari-
ty measure in the aspect of content-in-comment and to discover potential preference
20
of users [2]. Specifically, traditional information retrieval (IR) technology can be
used to analyze the semantics of content. To examine the semantic similarity among
posts, we use CKIP Chinese word segmentation system to parse and stem the crawled
contents and apply the analysis of frequency–inverse document frequency (TF-IDF)
weight to measure how important a word is to a document in a collection or corpus.
()
,
,
,
max
ij
ij
llj
freq
tf
freq
=
(5)
where
,
f
req
ij
as the raw frequency of term i appear in post j and
()
,
max
llj
f
req
is the number of times the most frequent index term, l, appears in post j . The inverse
document frequency for term i is formulated as
log
i
i
N
idf
n
=
(6)
where N is the total number of posts and
i
n
is the number of posts in which the
term i appears. The relative importance of term i to post j can be obtained by calculat-
ing
Then, we measure the similarity degree between people by a cosine similarity metric.
The similarity of corpus between user i and friend j
is defined as:
()
()
,cos,
||||
ij
Sim i j i j
CC
ij
==
G
G
G
G
G
G
(7)
where i
G
and j
G
are the two vectors in the m dimensional user space which is the
keywords to a person in a collection or corpus.
Finally, the total social similarity (SS) score is the sum of “friend-in-common” and
“content-in-comment” value.
Social Interaction. Social interaction is different from social similarity since social
interaction explicitly catches the factor of dynamic actions between people. It can be
used to evaluate the intimacy between two users. For instance, it is common for a user
to respond or forward someone’s message. It is reasonable to assume that more inte-
raction activities would lead to a higher probability to transmit the information as
they are usually interested in mutual messages. Given two users i,j ,social interaction
between them can be formulated as:
()
()
()
()
,,
||||
(, )
||||
response i j forward i j
response i forward i
SI i j
ΦΦ
=+
ΦΦ
(8)
()
,response i j
Φ is the set of responses generated by user j to user i’s posts and
()
,
f
orward i j
Φ is the set of forwards conducted by user j to user i’s posts.
Socail Activeness. Social activeness is used to calculate the activity intensity of a
user. A user with higher activeness indicates a lager level the user is engaged in in-
formation sharing or discussion with others and a higher probability to transmit the
21
information. We calculate the activeness of a user by the count of post records during
a period of time in a social platform. The formula is defined as below:
(,)
1
()
T
messages i t
i
SA i
T
=
=
Φ
(9)
where
(,)messages i t
Φ
is the total number of messages posted by user i at time period t.
The propagation strength measurement is used to evaluate the user whose network
propagation capability of a user. The propagation strength of a user is measured by
aggregating the propagation strength and can be computed in a recursive way. To
enhance advertising efficiently, it is important to pay attention on next layer’s propa-
gation capability. Though advertisers delivered advertisements to a social endorser
with high propagation capability, they can’t ensure the social endorser’s friends with
high propagation capabilities equally. Therefore, we thought friends’ propagation
capabilities would affect a social endorser’s propagation capability. Friends’ propaga-
tion capabilities became a dimension to measure a social endorser propagation capa-
bility. In other words, individual’s propagation capability is affected by their friends.
The propagation strength is formulated as below:
(
)
()
() () ( ) (, ) (, )
jFi
PA i SA i PA j SS i j SI i j
=+ +
(10)
where
()
F
i denotes a set of user i’s friends.
4 Experiments
Micro-blogging service has become one of the top tools for social media marketing.
Compared to traditional blogging, micro-blogging allows users to publish brief mes-
sages make people easy to read and repost.These characteristics: brief messages,
instant, easy to read and easy to share make micro-blogging become a good platform
to conduct social media marketing Therefore, in this research, we apply and validate
our proposed mechanism in micro-blogging systems. We conduct our experiments in
Plurk.According to Alexa, 2010, the user of Plurk is more than Twitter and 34.4% of
Plurk's traffic comes from Taiwan. Users of Plurk can connect with their friends via
lots of functions such as updating instant messages, sharing image or video to your
friends and responding friends’ messages. Besides the well-constructed network
structure, another important reason of choosing Plurk as our platform is they provide
the APIs for developers to easily request the data of users and networks which is
helpful to crawl more complete data to conduct our experimental work.
4.1 Data Description
In our mechanism, AdPlurker will send private messages with ads to users whom we
discover by different approaches. Users who receive the messages, which include
22
brief information of the ads and a recommended list of users who are also interested
in the ads and have higher propagation capabilities in their network. User can click
the hyperlinks of brief information to get detailed information and the click-through
record will be collected for evaluation work. Also, users can share the messages with
their friends who are recommended by the system. The transmitted message records
will be collected. To preventing click fraud, we recorded one click for each individual
user for the same advertisement. We conduct the experiment during the period of 11
April to 13 May.
In the preference module, we collect target users’ explicit preferences by ques-
tionnaires. Besides, in order to better realize users’ preferences, we collect implicit
preferences from the behaviors in Plurk. In the Plurk, “become fans” is a function for
users to follow others’ plurks and also declare their preferences for information type.
We use these data to match with the hierarchy of product category of Rakuten, one of
the famous online shopping mall, to structure the preference category tree of each
user. In the influence module, the out-degree measure and social popularity are taken
into consider. The friend links usually are the strongest links and imply the structural
influential in the network. A user is attention-getting since his/her plurks is often
replurked and responsed by others. It also means he/she is influential in content. In
propagation strength measurement, we analyze users’ occurring activities during the
recent six months: daily plurks, responses and replurks as the active and social inte-
raction measure, the similarity in friends and content as the social similarity measure.
We calculate the statistics of these as the propagation strength measurement.
We develop a Plurk robot named as AdPlurker and invite users who are active and
have used Plurk for a long time to join the experiments. Until April 2009, There are
107 users (55% male, 45% female) aging between 20-50. To simulate a real network
structure, our target users are formed with different locations and careers. There are
121,837 users and 971,014 plurks in our database. We collect data from the target
users to 3rd-4rd layers, due to the degrees of separation is limited the layer to 3rd-4rd
layer[10], and crawl their plurks, responses, and interactions with friends that hap-
pened within six months.
4.2 Experimental Results
In order to evaluate the performance of our proposed mechanism, we used the click-
through rate (CTR) [20] and repost-through rate (RTR) [13]. The former is a practical
statistics about advertising effectiveness; the latter is an effective means to evaluate
the eventual spread of the advertisement. Also, the two performance indicators are the
key measures in promoted tweets which is newly social advertisement platform pro-
posed by Twitter. The CTR formula is defined as:
clicks
CTR
delivered
Φ
=
Φ
(11)
where
clicksΦ
is is the total number of clicks and
delivered
Φ
is total number of ads
delivered. The RTR formula is defined as:
23
repost
RTR
delivered
Φ
=
Φ
(12)
where
repostΦ is the total number of repost and delivered
Φ
is total number of ads
delivered. We compare four online advertising approaches, which are commonly used
in micro-blogging. These different approaches are described as follows.
In-degree. It is the most common measure used to evaluate the influence of micro
blogging by the number of fans. This measure is currently employed by many other
third-part services, such as twiiterholic.com and wefollow.com
Ratio-degree. The measure is similar to the ratio between the number of a user’s
followers and the number of other people that the user follows. It was proposed from
the Web Ecology Project, an interdisciplinary research group based in Boston, Mas-
sachusetts.
Preference + Out-degree. Discovering the topic-influential nodes for delivering
advertising message by taking the advantage of the target advertising and social in-
fluence.
Social Diffusion. The approach we proposed in this study. We applied analytic hie-
rarchy process (AHP) to realize the final weight combinations of three components.
Figure 2 shows the results of different advertising strategies. According to the da-
tabase, the advertisements of in-degree approach got total 0.157 CTR in 2042 deli-
very times. Ratio-degree approach got 0.176 CTR in 3922 deliveries. The hybrid
approach of preference and out-degree got 0.217 CTR in 4596 deliveries. Our social
diffusion mechanism got 0.299 CTR in 7356 deliveries. Comparing the diffusion
performance in the four advertising approaches, we can observe that our proposed
social diffusion approach has the best coverage and exposure in advertising cam-
paign.
Fig. 2. Performance comparisons of various endorser discovering strategies.
5 Conclusions
In this paper, we propose a social diffusion mechanism to discover the nodes with the
strong propagation capability in delivering advertising information and recommend
each intermediate node a list of nodes with the high prior propagation so as to en-
hance the efficiency and effectiveness of spreading advertising message. We combine
the static factor, which includes individual preference and link structure of relation-
24
ship and the dynamic factor, which includes social interactions and social similarity
between the nodes, to develop our model. Our experimental results get positive out-
comes in both click-through rate and repost rate, and reveal some implicit connec-
tions between the components in the framework. A better CTR reflects that our me-
chanism can raise the visibility of advertising information. And a higher RTR indi-
cates a higher exposure of the advertising and reveals that users are interested in the
advertisement shared by friends and willing to share them with others. Our proposed
mechanism can widely extend the diffusion coverage of ads. It provides the advertis-
ing sponsors a powerful vehicle to successfully conduct advertising diffusion cam-
paigns.
References
1. Albadvi, A., Shahbazi, M. A hybrid recommendation technique based on product category
attributes. Expert Syst. Appl.(2009).
2. Berendt, B., Navigli, R. Finding your way through blogspace:Using semantics for cross-
domain blog analysis. 2006.
3. Delre, S. A., Jager, W., Bijmolt, T. H. A. and Janssen, M. A. Will it spread or not? The
effects of social influences and network topology on innovation diffusion. Journal of Prod-
uct Innovation Manageme(2010).
4. Gilbert, E., Karahalios, K. Predicting tie strength with social media. In Proceedings of the
Proceedings of the 27th international conference on Human factors in computing sys-
tems(2009) ACM.
5. Iribarren, J. L., Moro, E. Information diffusion epidemics in social networks. Physics and
Society(2007).
6. Kaplan, A. M., Haenlein, M. Users of the world, unite! The challenges and opportunities of
Social Media. Business Horizons 59-68.
7. Kempe, a., Kleinberg, J.,Tardos, É. Influential Nodes in a Diffusion Model for Social
Networks. Springer Verlag2005.
8. Kimura, M., Saito, K., Nakano, R., Motoda, H. Finding Influential Nodes in a Social Net-
work from Information Diffusio. Springer US 2009.
9. Leung, C. W.-k., Chan, S. C.-f.,Chung, F.-l. A collaborative filtering framework based on
fuzzy association rules and multiple-level similarity. Knowl. Inf. Syst.(2006) 357-381.
10. Li, Y.-M., Chen, C.-W. A synthetical approach for blog recommendation: Combining trust,
social relation, and semantic analysis. Expert Systems with Applications(2009) 6536-6547.
11. Li, Y.-M., Lien, N.-J. An endorser discovering mechanism for social advertising. In Pro-
ceedings of the Proceedings of the 11th International Conference on Electronic Com-
merce(2009) ACM.
12. Stelzner, M. Social Media Marketing Industry Report. 2009.
13. Twitter Promoted Tweets. 2010.
14. Wen, C., Tan, B. C. Y. and Chang, K. T.-T. Advertising Effectiveness on Social Network
Sites: An Investigation of Tie Strength, Endorser Expertise and Product Type on Consumer
Purchase Intention. 2009.
25