Identifying the k Best Targets for an Advertisement Campaign via Online
Social Networks
Mariella Bonomo, Armando La Placa and Simona E. Rombo
Department of Mathematics and Computer Science, University of Palermo, Palermo, Italy
Keywords:
Online Social Networks, Social Advertising, tf-idf, Profile Matching.
Abstract:
We propose a novel approach for the recommendation of possible customers (users) to advertisers (e.g., brands)
based on two main aspects: (i) the comparison between On-line Social Network profiles, and (ii) neighborhood
analysis on the On-line Social Network. Profile matching between users and brands is considered based on
bag-of-words representation of textual contents coming from the social media, and measures such as the Term
Frequency-Inverse Document Frequency are used in order to characterize the importance of words in the
comparison. The approach has been implemented relying on Big Data Technologies, allowing this way the
efficient analysis of very large Online Social Networks. Results on real datasets show that the combination
of profile matching and neighborhood analysis is successful in identifying the most suitable set of users to be
used as target for a given advertisement campaign.
1 INTRODUCTION
Social media have gained growing popularity in the
last few years, especially On-line Social Networks
(OSNs) that enable people to introduce themselves,
discuss about their preferred topics, establish and
maintain social connections. This leads advertisers
to invest more effort into communicating with con-
sumers through on-line social networking, that pro-
vides suitable platforms for advertising and market-
ing. An important issue in this context is how to
optimize the effects of marketing communication, by
taking advantage of the opportunities offered by the
OSNs.
In particular, advertisers aim to involve in their
campaigns those potential consumers who are the
most likely interested ones and, hopefully, could
spread the received advertisements to other interested
users. Automatic systems able to suggest a set of
target users for advertising campaigns provide three
main benefits: (i) minimization of costs for the dis-
semination of the advertising campaign through so-
cial media, which is often very expensive; (ii) im-
provement of the user experience in OSNs, since
only the possibly interested customers are contacted
with advertisements which could be useful for them;
(iii) avoid the spread of unuseful information through
OSNs.
Here we propose a novel approach for the rec-
ommendation of the k best possible consumers to be
suggested as target for a specific advertisement cam-
paign. The recommendation is based, on one hand,
on the comparison between the OSN profiles asso-
ciated to users (possible customers) and advertisers
(e.g., brands), according to the considered campaign.
On the other hand, also the chance that a specific user
may distribute the received advertisement to other in-
terested users is considered. In particular, bag of
words are used to represent user profiles and profile
matching is applied relying on the Term Frequency-
Inverse Document Frequency (TF-IDF), in order to
weight the importance of the words inside the text as-
sociated to user profiles. Moreover, for all users of
the considered OSN, their neighbors and correspond-
ing profiles are taken into account as well, in order
to understand to which extent it is convenient sending
the advertising to them.
We have applied our approach to real datasets,
which construction is part of the contributions pre-
sented here. Indeed, OSN datasets exist which are
publicly available but they usually include only net-
work topology, without extensive information on user
interests and other related information. On the other
hand, complementing the available network topolo-
gies via web-scraping starting from personal access
points is not trivial and often not possible. We present
a methodology for the association of contents to the
nodes of a OSN, given its topology, based on follow-
Bonomo, M., La Placa, A. and Rombo, S.
Identifying the k Best Targets for an Advertisement Campaign via Online Social Networks.
DOI: 10.5220/0010109201930201
In Proceedings of the 12th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2020) - Volume 1: KDIR, pages 193-201
ISBN: 978-989-758-474-9
Copyright
c
2020 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
193
ing the cross-linked references of public web-pages.
This allows to fulfil that neighbor nodes in the net-
work may share common contents more likely than
nodes very far each other. The obtained results are
promising, indeed the approach allows to correctly
identify the most suitable users in all the considered
situations.
2 RELATED WORK
Modeling the user profiles from social media raw data
is usually a challenging task. The approaches pro-
posed in the Literature to this aim may be roughly
classified in two main categories. The first category
includes approaches based on the analysis of user
generated contents (here referred to as semantic ap-
proaches). As for the approaches in the second cate-
gory, individuals are characterized by “actions”, e.g.,
visited web pages (action-based approaches). Our
framework belongs to the first category.
Semantic Approaches. The authors of (Schwartz
et al., 2013) use Differential Language Analysis
(DLA) in order to find language features across mil-
lions of Facebook messages that distinguish demo-
graphic and psychological attributes. They show that
their approach can yield additional insights (correla-
tions between personality and behavior as manifest
through language) and more information (as mea-
sured through predictive accuracy) than traditional a
priori word-category approaches.
The framework proposed in (Lin et al., 2014) re-
lies on a semi-supervised topic model to construct a
representation of an app’s version as a set of latent
topics from version metadata and textual descriptions.
The authors discriminate the topics based on genre in-
formation and weight them on a per-user basis, in or-
der to generate a version-sensitive ranked list of apps
for a target user.
In (Liang et al., 2018) the authors propose a dy-
namic user and word embedding algorithm that can
jointly and dynamically model user and word repre-
sentations in the same semantic space. They con-
sider the context of streams of documents in Twit-
ter, and propose a scalable black-box variational in-
ference algorithm to infer the dynamic embeddings of
both users and words in streams. They also propose a
streaming keyword diversification model to diversify
top-K keywords for characterizing users’ profiles over
time.
The first technique applied to brand-affinity
matching that is not an action-based approach has
been presented in (Bonomo et al., 2019). In partic-
ular, the authors present a profile-matching technique
based on tree-representation of user profiles and ap-
ply it on Facebook ego-networks. The approach pre-
sented here extends those results, showing that a suit-
able combination of profile-matching and neighbor-
hood analysis is more successful in identifying the
best k users for advertisements distribution.
Action-based Approaches. In (Provost et al.,
2009) individuals are associated each other due to
some actions they share (e.g., they have visited the
same web pages). The proximity between individuals
on networks built upon such relationships is informa-
tive about their profile matching. In particular, brand-
affinity audiences are built by selecting the social-
network neighbors of existing brand actors, identified
via co-visitation of social-networking pages. This is
achieved without saving any information about the
identities of the browsers or content of the social-
network pages, thus allowing for user anonymization.
In (Ahmed et al., 2011) compact and effective user
profiles are generated from the history of user actions,
i.e., a mixture of user interests over a period of time.
The authors propose a streaming, distributed infer-
ence algorithm which is able to handle tens of mil-
lions of users. They show that their model contributes
towards improved behavioral targeting of display ad-
vertising relative to baseline models that do not incor-
porate topical and/or temporal dependencies.
In (Iglesias et al., 2012) a computer user behav-
ior is represented as the sequence of the commands
she/he types during her/his work. This sequence
is transformed into a distribution of relevant subse-
quences of commands in order to find out a profile
that defines its behavior. Also, because a user profile
is not necessarily fixed but rather it evolves/changes,
the authors propose an evolving method to keep up to
date the created profiles using an Evolving Systems
approach.
The observation that behavior of users is highly
influenced by the behavior of their neighbors or com-
munity members is used in (Xie et al., 2014) to enrich
user profiles, based on latent user communities in col-
laborative tagging.
3 PROPOSED APPROACH
The main goal of the proposed approach is to identify
the most suitable k possible buyers to whom distribut-
ing a given advertisement campaign. To this aim, two
important aspects have to be taken into account:
KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval
194
Ideally, users to whom distributing the campaign
should have interests compatible with the specific
features of the advertiser (i.e., the brand).
It would be better if the chosen possible buyers
would know other users whose interests are close
to those expected for the campaign success as
well.
The second point has a twofold effect. Indeed, it
can be useful in order to obtain the next k users to con-
tact and, at the same time, maximize the chance that
they can be contacted directly by the buyers selected
at the current step.
In order to accomplish the two above points, we
have based the presented research on the use of Online
Social Networks.
Online Social Network (OSN). We represent an
Online Social Network as an undirect graph N =
(V, E) such that nodes in V are associated to the users
and two nodes are linked in N if a social relationship
(e.g., friendship, common interests, etc.) between
them occurs in the represented OSN. In addition to
the topological representation of an OSN, further de-
tails are necessary in order to characterize each node.
User Profile. User profiles complement network
topology information. In particular, each node in the
network points to data associated to a user and re-
trieved from the considered social media. What is im-
portant for this research, is textual information about
user general interests and activities, coming for exam-
ple from private communications, posts, comments,
short text messages. Therefore, the user profile of u
is represented here by a text T
u
characterizing u with
references to the considered OSN.
Brand Profile. Also a brand profile is represented
by a text, that can be for example easily extracted
from the web-page describing brand activities or from
other textual documents containing information on
the advertisement campaign. In the following, we
refer indistinctly to brand profile and advertisement
campaign, since both may be described by textual
documents and then handled in the same way in the
context of the proposed approach.
3.1 Profile Matching
Let u be a node in an input OSN N and T
u
be its
user profile. Moreover, let T
b
the text associated to
the brand profile. The first step of our approach is
to understand how much T
u
and T
b
are “similar”, i.e.,
to which extent they match each other. To this aim,
we consider TF-IDF and cosine similarity measures
in order to understand if and how much textual con-
tents associated to u and to the brand are semantically
related. This is sketched in the following, for the spe-
cific case under consideration (i.e., only two textual
documents, T
u
and T
b
).
Importance of Words. Let w
i j
be a word occurring
in the text T
j
( j = 1, . . . , m). The TF-IDF function for
w
i j
is defined as:
T F-IDF(w
i j
) = T F(w
i j
) IDF(w
i j
)
such that:
T F(w
i j
) =
|w
i j
|
|T
j
|
where |w
i j
| is the frequency of the term w
i j
in the
text T
j
and |T
j
| is the number of words in T
j
, and:
IDF(w
i j
) = log
m
h
, where h m is the number of texts
where w
i j
occurs.
Affinity between Profiles. Let V
u
and V
b
be two ar-
rays of k real values. The cosine similarity between
V
u
and V
b
is defined as:
C S (V
u
, V
b
) =
k
i=1
V
u
[i] V
b
[i]
q
k
i=1
V
u
[i]
2
q
k
i=1
V
b
[i]
2
The affinity between the profiles associated to an user
and a brand is then computed as the cosine similarity
between arrays containing the TF-IDF values of the
words occurring in T
u
and T
b
:
A(T
u
, T
b
) = C S (V
u
, V
b
)
3.2 Neighborhood Analysis
In order to make more effective the advertisement
campaign, for each node u in V , it is important not
only to measure to what extent its profile matches
with the brand profile, but also how many nodes in the
neighborhood of u could be possibly interested in that
campaign as well. That is, the best targets are those
nodes which profile matches with the brand, and that
are surrounded by other nodes with this same feature.
Node Centrality. Let u be a node in the set of ver-
tices V and T
u
and T
b
be the user and brand profiles,
respectively. Moreover, let N
u
be the set of nodes
linked to u by at least one edge in the set of edges
E of N . Then, the centrality of u for the given con-
sidered brand (or advertisement campaign) is defined
as:
Identifying the k Best Targets for an Advertisement Campaign via Online Social Networks
195
C (u, b) =
vN
u
A(T
v
, T
b
)
|N
u
|
It is worth pointing out that, in order to focus the ad-
vertising campaign on those interested users only, a
threshold value can be chosen on the affinity values
according to which filtering only nodes in the network
scoring affinity values larger than that threshold.
Node Utility. As already explained, the final aim of
our approach is to identify the best k nodes to which
distribute advertisements according to their profile
matching with the brand (or campaign, respectively).
On the other hand, in order to maximize the gain, we
are also interested into detecting nodes which neigh-
bors in the OSN may be interested in the same adver-
tisements. To this respect, the utility of a node for a
specific brand/campaign is defined as follows:
U(u, b) = α · A(T
u
, T
b
) + (1 α) · C (u, b)
where α is a real value in [0, 1] used to balance two
different contributions, i.e., the match between user
and brand, and the match between user neighbors and
brand.
3.2.1 Example
Figure 1 depicts a small OSN and, for each node, the
corresponding affinity value that has supposed to be
computed with respect to a given brand is also shown.
Suppose that the brand is interested to send its ad-
vertising campaign to 5 nodes on this network (i.e.,
k = 5).
Figure 1: A small OSN. For each node, the corresponding
affinity value is also shown.
As an example, for Node 1:
C (1, b) =
0.7 + 0.8 + 0.8 + 0.5 + 0.8
5
= 0.72
and, for α = 0.4:
U(1, b) = 0.4 · 0.7 + (1 0.4) · 0.72 = 0.71
while for α = 0.6:
U(1, b) = 0.6 · 0.7 + (1 0.6) · 0.72 = 0.7
Tables 1-3 show the utility and centrality values for
nodes in the example OSN, for three different values
of α. In particular, nodes are sorted according to the
utility, and filtered such that only those with utility
values larger than 0.6 are shown in the tables for each
considered α.
Table 1: Top nodes for α = 0.4, sorted according to their
utility values.
Node utility centrality
3 0.74 0.7
2 0.73 0.75
7 0.72 0.67
9 0.72 0.8
1 0.71 0.72
Table 2: Top nodes for α = 0.5, sorted according to their
utility values.
Node utility centrality
3 0.75 0.7
7 0.73 0.67
2 0.72 0.75
8 0.72 0.75
4 0.7 0.6
Table 3: Top nodes for α = 0.6, sorted according to their
utility values.
Node utility centrality
3 0.76 0.7
7 0.74 0.67
2 0.72 0.75
4 0.72 0.6
8 0.72 0.75
From Tables 1-3, it is evident that Node 3 is the best
target for the input brand for all considered α val-
ues, according to the supposed affinity values. This
is due to the fact that Node 3 and its two neighbors
1 and 2 all score good affinity values, therefore Node
3 keeps its top ranking position both when it is given
higher importance to its affinity with the brand (larger
α value) and when, instead, the focus is on the affin-
ity of its neighbors (smaller α value). In the transi-
tion from smaller to larger α values, the configuration
of top ranking change for the further four positions.
From α = 0.4 to α = 0.5, Nodes 2 and 7 exchange
their position in the ranking, due to the fact that both
them have neighbors with high affinity values but the
affinity between Node 7 and the brand is higher than
that of Node 2. Another effect is that Nodes 1 and 9
KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval
196
are replaced by Nodes 4 and 8, respectively. Again,
all such nodes have neighbors with high affinity val-
ues, but the latter nodes have larger affinity values
than the former ones. Analogous considerations can
be done from α = 0.5 to α = 0.6, where the only effect
is the inversion between Nodes 4 and 8 in the ranking.
4 RESULTS
The main goal of our experimental validation has
been to verify on large OSN datasets if the choice of k
best targets according to the measures introduced here
is effective. To this aim, a first important problem to
be solved has been the construction of the input OSN.
Indeed, while a number of social network graphs are
publicly available, the same is not true for network
users profiles. In the following of this section, we
first discuss these aspects related to OSN construc-
tion, and then present some results we have obtained
by applying our approach on datasets coming from
the real world. The proposed approach has been im-
plemented in Java 1.8 under Apache Spark 1.6. To
this respect, the use of Big Data Technologies allow
to exploit the software tool also on very large OSNs.
4.1 Network Construction
OSN graphs are available for example from Standford
website (https://snap.stanford.edu/data/).
We have considered the twitter-2010 OSN from
that repository, having 90, 908 vertices and 443, 399
edges. Unfortunately, the available OSNs consist
only on the graph topology, no information about
user interests and profiles are publicly available. Web
scraping has been used here in order to collect and
extract useful contents for user profiles characteri-
zation. In particular, we have avoided to associate
randomly the information obtained by web scraping
to nodes in the considered OSN graph, due to the
fact that a random association would have altered
the natural mechanism according to which users in
the same neighbors have similar interests. In order
to mimic such a mechanism, which is important for
our approach (indeed the introduced measures aim at
detecting neighbor nodes with similar interests), we
have proceeded as follows.
We have first randomly selected 20 seed nodes
from the twitter-2010 OSN and 20 web-pages fo-
cused on different topics (cooking, fashion, cars, etc.).
Indeed, with a certain margin of simplification, we
have assumed that a user profile may be obtained
by scraping the contents of a web-page on a specific
topic. Then, a visit in depth of the OSN has been per-
formed starting from each of the seeds and stopping
when the entire network was visited. For each new
node to be visited, a new web-page has been visited
as well, following the cross-page links on the consid-
ered web-pages.
4.2 Experimental Validation
Our experimental analysis has been devoted to under-
stand to what extent our approach is effective, in or-
der to identify the k most convenient nodes in the in-
put OSN to which distribute the advertisement. As
already explained, the main aim here is to optimize
two different aspects when identifying the best tar-
gets, that is, the fact that interests of considered users
are related to the campaign contents, and the fact that
they have “friends” on the OSN potentially interested
to the distributed advertisements. We have considered
the web-pages associated to four brands, listed in Ta-
ble 4.
Table 4: The considered brands and their associated web-
pages.
Brand Web-page
AlphaRomeo www.alfaromeo.it
Amarelli www.amarelli.it
Carpisa www.carpisa.it
KikoCosmetic www.kikocosmetics.com
We have considered the OSN constructed as de-
scribed in the previous section and we have computed,
for each of the four brands, the different values of
affinity and utility (with α = 0, 25;0, 5;0, 75) for all
nodes in the network. Then, we have ranked them
in descending order, according to each of these mea-
sures. We have supposed that the number of target
nodes is k = 100 and we have fixed to 0.6 the mini-
mum value of affinity between user and brand profiles
in order a user to be considered a possible target.
The obtained results have been compared with a
random choice of the k nodes to which distribute the
advertisement. For 100 different times, 100 nodes
have been extracted from the set of vertices V and
the affinity between their and brand profiles have been
computed at each time. The obtained results for
the different brands do not present significant differ-
ences, therefore we illustrate only those regarding Al-
phaRomeo in Table 5. In particular, the considered
method is specified in the first column of the table,
and for the Random generation we have considered
the average of obtained results. For each method, the
number of nodes presenting an affinity value larger
than the chosen threshold when the first k nodes in
the corresponding ranking is chosen is shown in the
third column. It is interesting to observe that, with
Identifying the k Best Targets for an Advertisement Campaign via Online Social Networks
197
respect to the random choice, both Affinity and Util-
ity with a high value of α (0.75) improves of one or-
der of magnitude. Indeed, in this two latter cases, all
the considered nodes have affinity values above the
threshold. This shows that the profile matching at the
basis of our approach is effective in the selection of
target users for an advertising campaign. However,
the second aspect to take into consideration is related
to the number of possible further interested users that
can be reached by the advertisement, starting from
those k. To this respect, the last column of Table 5
shows how many distinct nodes are in the neighbor-
hoods of the first k ones (according to the ranking ob-
tained for each method). The second column of the
table shows the total number of nodes with affinity
values larger than the threshold that can be reached
starting from the first k, for each ranking. It is evi-
dent that, again, the worst performance is obtained by
the Random method, whereas the best one by Util-
ity with α = 0.5 in this case. This confirms what ex-
pected, that is, neighborhood analysis associated to
profile matching is the most promising choice.
Figures 2-5 provide a graphical illustration of the
links between the first 10 target nodes for each of the
considered brand according to the method Utility with
α = 0.5. In particular, the web-page of the brand is the
central node, and the web-pages associated to the first
10 nodes in the ranking are depicted around, showing
also the existing links among them in the OSN. For all
brands, most of the considered target nodes are con-
nected in paths, trees or small communities.
Figure 2: Links among the first 10 target nodes for Alpha
Romeo.
Tables 6-9 show the web-links of the top 10 nodes
for each brand, and their values of utility and affin-
ity. Moreover, in the last column it is also reported,
for each node, the number of neighbors that are target
nodes as well (i.e., their affinity value is larger than
the considered threshold). It is evident that the top
target nodes refers to web-pages which contents are
strictly related with those of the brand, in each case.
Figure 3: Links among the first 10 target nodes for
Amarelli.
Figure 4: Links among the first 10 target nodes for Carpisa.
Figure 5: Links among the first 10 target nodes for Kiko
Cosmetics.
KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval
198
Table 5: Total number of nodes (second column) with affinity values larger than the chosen threshold identified by each
method (first column), fraction of target nodes directly reached (third column) or instead detected from the neighborhoods
(fourth column).
Method # of Target Nodes Directly Reached From Neighborhoods
Affinity 184 100 84
Utility (α = 0.25) 152 64 88
Utility (α = 0.5) 192 99 93
Utility (α = 0.75) 181 100 81
Random 99 13 86
5 CONCLUDING REMARKS
We have discussed here how the combination of in-
formation retrieval measures for profile matching and
neighborhood exploration in OSNs may be successful
in order to identify a set of target users for the distri-
bution of advertisements. In particular, such users not
only have interests related to the contents of the adver-
tisement, but may also potentially spread the received
advertisements to other interested users in the OSN.
This allows to minimize costs for advertising cam-
paigns, improve user experience in OSNs and avoid
spread of unuseful information through OSNs.
Results obtained by the measures introduced here
on real datasets are promising. However, we are con-
scious that the proposed approach relies on a naive,
although effective, technique for neighborhood ex-
ploration. In our future work we plan to extend it
by taking into account more complex node centrality
measures (Giancarlo et al., 2019; Mohammed et al.,
2020). Moreover, we will explore the direction of
including OSNs community detection (Wadhwa and
Bhatia, 2014) in our analysis, in order to identify
compact groups of nodes with interests related to the
considered input advertising campaigns.
Finally, we conclude with the following observa-
tion. An important problem in the context of OSNs
analysis is the absence of publicly available datasets
including not only network topology, but also struc-
tured information related to the network users, such
as interests, general data, actions, etc.. It is worth
to point out that the construction of such datasets
via web-scraping starting from personal access points
on the OSN presents several problems, among which
data privacy constraints, the fact that the obtained
networks would be mostly ego-networks (Arnaboldi
et al., 2017; Kwon et al., 2019), and the difficulty in
building networks that reflect the sizes of real OSNs,
often very large (Peng et al., 2017; Wu et al., 2020).
Therefore, providing suitable OSN public datasets
which contain both topological and semantic data
would be a valuable contribution for the scientific
community. We plan to extend in this direction the
procedure described here for the construction of big
OSNs, and to provide a public repository containing
such datasets.
ACKNOWLEDGEMENTS
Part of the research presented here has been funded
by the MIUR-PRIN research project “Multicrite-
ria Data Structures and Algorithms: from com-
pressed to learned indexes, and beyond”, grant n.
2017WR7SHH, and by the INdAM - GNCS Project
2020 Algorithms, Methods and Software Tools for
Knowledge Discovery in the Context of Precision
Medicine”.
REFERENCES
Ahmed, A., Low, Y., Aly, M., Josifovski, V., and Smola,
A. J. (2011). Scalable distributed inference of dy-
namic user interests for behavioral targeting. In Pro-
ceedings of the 17th ACM SIGKDD International
Conference on Knowledge Discovery and Data Min-
ing, KDD ’11, pages 114–122. ACM.
Arnaboldi, V., Conti, M., Passarella, A., and Dunbar, R. I.
(2017). Online social networks and information dif-
fusion: The role of ego networks. Online Social Net-
works and Media, 1:44–55.
Bonomo, M., Ciaccio, G., Salve, A. D., and Rombo, S. E.
(2019). Customer recommendation based on profile
matching and customized campaigns in on-line social
networks. In ASONAM’19: International Conference
on Advances in Social Networks Analysis and Mining,
Vancouver, British Columbia, Canada, 27-30 August,
2019, pages 1155–1159.
Giancarlo, R., Greco, D., Landolina, F., and Rombo,
S. E. (2019). Network centralities and node rank-
ing. In Ranganathan, S., Gribskov, M., Nakai, K., and
Sch
¨
onbach, C., editors, Encyclopedia of Bioinformat-
ics and Computational Biology, pages 950–957. Aca-
demic Press, Oxford.
Identifying the k Best Targets for an Advertisement Campaign via Online Social Networks
199
Iglesias, J. A., Angelov, P., Ledezma, A., and Sanchis, A.
(2012). Creating evolving user behavior profiles auto-
matically. IEEE Transactions on Knowledge and Data
Engineering, 24(5):854–867.
Kwon, Y. D., Mogavi, R. H., ul Haq, E., et al. (2019).
Effects of ego networks and communities on self-
disclosure in an online social network. In ASONAM
’19: International Conference on Advances in Social
Networks Analysis and Mining, Vancouver, British
Columbia, Canada, 27-30 August, 2019, pages 17–24.
Liang, S., Zhang, X., Ren, Z., and Kanoulas, E. (2018).
Dynamic embeddings for user profiling in twitter. In
Proceedings of the 24th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining,
KDD 2018, London, UK, August 19-23, 2018, pages
1764–1773.
Lin, J., Sugiyama, K., Kan, M.-Y., and Chua, T.-S. (2014).
New and improved: Modeling versions to improve
app recommendation. In Proceedings of the 37th
International ACM SIGIR Conference on Research
& Development in Information Retrieval, SIGIR
’14, pages 647–656. ACM.
Mohammed, A., Zhu, F., Sameh, A., et al. (2020). Efficient
algorithms based on centrality measures for identifica-
tion of top-k influential users in social networks. Inf.
Sci., 527:88–107.
Peng, S., Wang, G., and Xie, D. (2017). Social influence
analysis in social networking big data: Opportunities
and challenges. IEEE Network, 31(1):11–17.
Provost, F., Dalessandro, B., Hook, R., Zhang, X., and Mur-
ray, A. (2009). Audience selection for on-line brand
advertising: Privacy-friendly social network targeting.
In Proceedings of the 15th ACM SIGKDD Interna-
tional Conference on Knowledge Discovery and Data
Mining, KDD’09, pages 707–716, New York, NY,
USA. ACM.
Schwartz, H., Eichstaedt, J., Kern, M., et al. (2013). Per-
sonality, gender, and age in the language of social
media: The open-vocabulary approach. PLoS ONE,
8(9):e73791.
Wadhwa, P. and Bhatia, M. P. S. (2014). Community detec-
tion approaches in real world networks: A survey and
classification. Int. J. Virtual Communities Soc. Netw.,
6(1):35–51.
Wu, Y., Huang, H., Wu, N., et al. (2020). An incentive-
based protection and recovery strategy for secure
big data in social networks. Information Sciences,
508:79–91.
Xie, H., Li, Q., Mao, X., Li, X., Cai, Y., and Rao, Y. (2014).
Community-aware user profile enrichment in folkson-
omy. Neural Networks, 58:111–121.
KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval
200
APPENDIX
Table 6: First 10 target nodes for Alpha Romeo.
Network’s node Brand: Alpha Romeo Utility Affinity # of Target Nodes
Node 1 blogmotori.com 0.84 0.64 5
Node 2 autoblog.it 0.80 0.61 2
Node 3 blogdimotori.it 0.88 0.87 5
Node 4 hdmotori.it/auto 0.74 0.71 2
Node 5 motori.news 0.73 0.69 2
Node 6 motoblog.it 0.91 0.67 3
Node 7 motoblogtrotter.com 0.91 0.71 4
Node 8 automobile-domani.blogspot.com 0.88 0.64 2
Node 9 autologia.net 0.80 0.62 3
Node 10 dotcar.it/ 0.94 0.92 3
Table 7: First 10 target nodes for Amarelli.
Network’s node Brand: Amarelli Utility Affinity # of Target Nodes
Node 1 lucake.it 0.89 0.78 2
Node 2 hovogliadidolce.it 0.93 0.86 4
Node 3 brindando.com 0.85 0.70 4
Node 4 cioccolatoeliquirizia.it 0.93 0.61 2
Node 5 caramelleonline.com/blog 0.75 0.79 2
Node 6 mentaeliquirizia.com 0.89 0.63 2
Node 7 veganblog.it 0.75 0.63 3
Node 8 cacaocrudo.it/it 0.87 0.69 3
Node 9 blog.cookaround.com/dolcevanilia 0.76 0.76 3
Node 10 blog.rigato.net/tag/liquirizia 0.92 0.70 3
Table 8: First 10 target nodes for Carpisa.
Network’s node Brand: Carpisa Utility Affinity # of Target Nodes
Node 1 concosalometto.com 0.75 0.74 1
Node 2 ireneccloset.com 0.71 0.61 4
Node 3 bagsandfruits.com/it/blog 0.82 0.67 2
Node 4 ilbellodelleborse.com/blog 0.75 0.70 2
Node 5 latolfetana.com/blog/page/2 0.77 0.66 2
Node 6 elle.com 0.79 0.61 2
Node 7 mondoborse.com 0.88 0.73 2
Node 8 bags.stylosophy.it 0.70 0.62 3
Node 9 saragiunti.it 0.92 0.90 3
Node 10 saragiunti.it/blog 0.75 0.68 2
Table 9: First 10 target nodes for Kiko Cosmetics.
Network’s node Brand: KikoCosmetic Utility Affinity # of Target Nodes
Node 1 blog.cliomakeup.com 0.81 0.70 3
Node 2 makeupdelight.com 0.62 0.61 2
Node 3 claudia-makeup.com/blog-trucco 0.82 0.73 2
Node 4 polveredistellemakeup.com 0.77 0.62 3
Node 5 aboutbeautymakeup.wordpress.com 0.85 0.68 4
Node 6 donnaedintorni.com/blog-makeup 0.85 0.80 3
Node 7 loscrigno.it/beautycase 0.77 0.70 4
Node 8 sabbioni.it/it/blog.php 0.85 0.71 2
Node 9 ilmiomakeup.it 0.76 0.64 3
Node 10 follettarosamakeup.com 0.75 0.63 2
Identifying the k Best Targets for an Advertisement Campaign via Online Social Networks
201