You are on page 1of 10

Proceedings of the 52nd Hawaii International Conference on System Sciences | 2019

A Social Network Analysis (SNA) Study On Data Breach Concerns Over


Social Media
Naga Vemprala
Glenn Dietrich
The University of Texas at
The University of Texas at
San Antonio
San Antonio
Naga.Vemprala@utsa.edu
gdietrich@utsa.edu

Abstract about security [4], [5], estimating the loss of market


value post breach [24], [25] and on the after effects of
In the current era of digital devices, the concerns data breaches by analyzing the responses of data breach
over data privacy and security breaches are rampant. victims [7] and by tagging the psychological theories to
Understanding these concerns by analyzing the the breach situation [6]. The earlier studies in the
messages posted on the social media from linguistic Information Systems space focused on analyzing the
perspective has been a challenge that is increasing in outcome of the data breach using event studies [24],
complexity as the number of social media sites increase [25] and processed the responses of victims by
and the volume of data increases. We investigate the surveying them using a set of well-phrased questions
diffusion characteristics of the information attributed around the situation. However, there is not much
to data breach messages, first based on the literary research available on the data breach
aspects of the message and second, we build a social information/discussion encompassing public
network of the users who are directly involved in reactions stated over social media. Analyzing the first-
spreading the messages. We found that the messages hand opinions from a wide range of victim-groups on
that involve the technicalities, threat and severity social media facilitates in bringing valuable insights
related security characteristics spread fast. Contrary about public reactions to the post- breach situation.
to conventional news channels related posts on social These insights help the organizations and lawmakers to
media that capture wide attention, breach information protect the reputation and mitigate the losses incurred
diffusion follows a different pattern. The messages are through the data breach. Specifically, in this study we
widely shared across the tech-savvy groups and people address the following research questions: RQ1: What
involved in security-related studies. Analyzing the characteristics of a breach message makes the
messages in both linguistic and visual perspective information diffuse widely across the social
through social networks, researchers can extract network?
grounded insights into these research questions. RQ2: Does the presence of evidence in the form of
media files, URLs and videos, drive information
diffusion?
1. Introduction RQ3: Who among the social media users exchange
wide amounts of breach messages?
Data breaches have been one of the most popular
To answer these questions, we collected the social
discussion topics over the last decade. There are
media messages about the data breaches from
various companies and many customers that are
Twitter.com. Twitter data is extensively used in
affected because of the data breaches at commercial
research because of its popularity and extensive
firms including BJ’s Stores, Target, Home Depot, the
reliance by various user groups [26]. We studied the
Office of Personnel and Management (OPM), Equifax,
prior literature on this subject to retrieve and analyze
Anthem, Coca-Cola are to name a few. The Pew
key aspects about the victim-groups. Based on the
research showcase finds that 90% of Americans feel
findings from earlier studies [27], data breach related
they have lost control of their personal data [1]. Many
keywords and the survey questions used, we have
scholarly studies were conducted around data breaches
retrieved 16 keywords that have been used in previous
in many different domain interests. Some of the studies
studies. To find the contextually similar words and to
focused on making policy changes, stricter laws [2],
extract the words for the study, the words were
estimating cost of data breaches [3], educating users
converted into a meaningful numeric form as a real-
valued vector and we represented them on a scale. For
this reason, we have used the word

URI: https://hdl.handle.net/10125/60155 I
SBN: 978-0-9981331-2-6
(CC BY-NC-ND 4.0)
Page 7186
embedding [19] (The vector representation of words exchanging information. Especially with the social
using real numbers) to represent semantically media becoming more and more available through
connected words together. The numerically extracted many channels, the concerns are equally rising at an
words that are contextually similar in meaning to the 16 unprecedented rate. Bulgurcu et.al. studied the
words we previously selected, allows us to form the phenomena of tightening security controls and
basis for building the features for our model. We used educating employees and users on the data security. [8]
the Negative Binomial Regression, a statistical method These measures improve the overall security within an
employed by the prior study to logically conclude organization and reduce the effect for a data breach.
which set of features significantly contributed to the Some of the researchers focused on the ethical aspects
information diffusion. The first two research questions of the breach and studied how the organizations should
of our study can be answered using the results from our consider ethical aspects of managing customer’s
statistical methods. Finally, we have segregated the private information [9]. These studies are more focused
users who shared the breach information on social towards identifying the root causes and how the future
media and built a Social Network of the users. This data breaches could be avoided. Victim-groups are
provides a visual representation in answering our third asked to respond to a set of pre-defined survey
research question. questions aligned in the direction of research questions.
These responses are further analyzed for understanding
From a technical and methodological standpoint, our the user perception towards capturing the user
study is novel and first of its kind building the social responses through a set of survey questions. These
network on breach message information. We provide surveys are not only costly but also poses the risk of
three important contributions through our study: building a non representational sample.
1) A highly accurate and broadly demographic Moreover, these studies suffer a high degree of
sample: Utilized social media data in real-time margin-of-error [20] due to various unaccounted
for capturing user responses to the data factors including users subject matter knowledge. This
breaches avoiding surveying users. happens even after considering heterogeneous user’s
2) Representation sample: Using word groups while collecting the survey data. To address the
embedding for extracting contextually similar issue of addressing representational sample,
words from the entire set of user generate Chakraborty et al., collected data using Qualtrics, from
messages not only looks at the meaning of the a random sample of user groups segregated by gender
words but also the context a word represents. and age. [21]. With the help of the tool provided by
This approach is more robust than using a Qualtrics, the responses for the survey questions were
linguistic dictionary as a reference to extract captured. One limitation of the study is about the
similar meaning words from the dictionary. concentration area of shopping experience of various
3) Quick analysis on user characteristics who users depending on the recent data breach trends. The
involve in message diffusion through social problem of limited sample size applicable to one
networks backs the evidence provided by domain area could be addressed if a diverse sample
statistical results on the dissemination with heterogeneous users is considered for the study.
characteristics of a message. This is only possible with the first-hand data collected
from social media. Social Media data is one such
platform where users communicate at will on various
2. Theoretical model and Hypothesis
topics. The research interests of users also cover vast
interest groups. Mining the responses of users on data
Our study broadly builds on the domain aspects of breach topics provided valuable information. In this
data breaches and systems security, two streams of study we collected the data breach related information
technical research from the existing literature in social from Twitter.com for mining the information, which
networking analysis and the sparse neural networks. provides an alternate solution of capturing survey
The sub-sections provide detailed explanation about responses and we extracted the user interests by
prior theoretical and methodological techniques utilized reading through their public profile pages.
in the prior studies and finally helps in building
hypothesis is derived based on the theoretical
2.2. Neural Networks (Word2Vec)
background.

2.1. Data Breach studies The first step of the research methodology is to find
the contextually similar words for the given set of input
keywords. We used the prior literature for building a
With the increased number of data breaches, there has word dataset that was as the input to query for the
been a huge uproar among all the user groups contextually and semantically similar words. This
allows us to build the features (message characteristics)

Page
71872
of data breach given input word, utilizing in research. However, mere count, then log-linear
messages. The the text corpus for training. multiple linear or regression is usually
contextually similar In our study, we used a set multivariate regression employed. In
words are used for of machine learning, analysis with Ordinary analyzing the
establishing the including the Skip-gram Least Squares (OLS) information diffusion
features for the model and data analytics method is not a one fit considering the
regression model, tools to dynamically learn solution for many sentiments carried in
which is explained in word embeddings from the problems. Especially when the message, Stieglitz
the next sub-section. ongoing supply of the dependent variable is a and Dang- Xuan [16]
Advancements in the fragmented social media continuous variable which utilized NBR with
linguistic studies messages. We performed in the context of Twitter, it retweets count as the
enabled scholars to text pre- processing to is the retweet dependent variable
retrieve the contextual filter the unrelated and sentiment,
words around a given information that induce the hashtags, and URL,
query word. noise contained in these and followers are
Converting words into messages. This eliminates independent
numeric format helps certain messages that are variables. A similar
in processing text and not related to the data analysis is carried out
representing it in a breach. E.g: Username by Lee, Agrawal and
space vector based on containing breach which is Rao [17] to estimate
the similarity. This used in a tweet. Some the message diffusion
process of retrieving announcement about on rumors during the
numeric word vector cybersecurity training, Boston Bombing. We
in low-dimension posts that contain taglines adapted the same
space is known as with “breach” related statistical method for
word embedding. keywords etc. were also capturing the
Word2Vec is a neural eliminated from the study. significant message
network based We created the dimensions characteristics as the
algorithm [10] that in tweets using shallow assumptions from
takes words as inputs (single layer network) previous studies holds
by converting the neural networks technique valid for our study.
words into word called Word2Vec, which Along with the URL,
embeddings. Using maps a word to the target the number of
these word (target is also a word). hashtags we have
embeddings that are Word2Vec focuses on used, the message
plotted with similar shared meaning of the characteristics like
words, a given input words rather than on links. stress, public opinion,
word can be queried For example, two words anxiety, efficacy,
to retrieve the are connected if their use threat, hacking related
contextually similar of their concepts overlap. It and company-specific
words around it. For is an unsupervised machine discussions have also
this the second learning model that does been as features. We
version of the not require any labels and ran the Negative
algorithm is used. can learn the weights of Binomial Regression
There are two words as word vector using SPSS software.
versions of representations
Word2Vec, incorporating most of the
Continuous Bag of semantically rich
Words (CBOW) [11] information. These vectors 2.4. Social Network Analysis
and Skip-gram model can then be used to find
[10]. CBOW takes a similarities between words Social network
set of words to get the to get relations. analysis (SNA) is a
missing contextual well-developed
word, while the skip- 2.3. Negative Binomial methodology that
gram model takes an Regression (NBR) uses network graphics
input word to fetch to illustrate social
the set of contextual structures among
Regression analysis is the
words around the related subjects. A
most popular method used
Page
71882
SNA constructs, a protecting the or a news, self
network variety of systems efficacy in addressing
graph properties such [22] [23]. The the problem, threat
consists of as centrality, major concerns and severity
nodes diversity, are broadly associated with the
(also cohesion, and related to breach due to
known as equivalence stress, anxiety, information hacking
vertices) can be opinion about and loss of privacy.
and measured and company image We have established
relations adopted to our
(also summarize or search keywords based on initial data we collected, it is
known as predict the the words related to these evident that majority of the
edges). evolution of the aspects. Using these profiles are interested in security,
The nodes network and keywords, we extract the technology, news, social aspects
can be the semantically similar words and other general interests. Based
individual performance of that are used within the on this foundation we have
or group individual entire set of messages. framed our hypothesis as below:
entities in entities in the These words help in H1: Stress, anxiety and opinion
a network Exploring identifying message about company should have
[12] while Untapped characteristics. Not all positive effect on the information
edges are Markets by characteristics of a message dissemination. H2: Evidence
the Tracing play the catalyst role in based messages (messages with
connectio Through Weak information propagation. URLs and more hashtags)
ns Links network. Identifying these keywords contribute more positively to
between In this study, that are significant in information dissemination
nodes and we use nodes to information propagation H3: Security and Technology
can be represent and visualizing groups of related interest groups creates a
either uni- similar group users based on their dense social network on
or bi- of users and the interests contributes to the information sharing and
directional communication information diffusion, is positively contributes for
(dyadic). between these very helpful. To build a information dissemination
These groups are network, we need to
notions represented establish a relationship Running a negative binomial
were using the between the aspects of the regression using the features we
initially edges. message. We use the prior framed helps in answering the
introduced literature studies and the first two hypotheses, where we
by 2.5. Hypothesis survey questions to frame obtain the coefficients for the
sociologis the initial set of words to regression model with the
ts and draw the semantically
Our research retweets as dependent variable on
subsequen related words used in the
plan consists of a logarithmic scale. If the
tly overall set of data breach
setting up a coefficients values are positive
broadly related message. The
dictionary of and large they contribute more
applied to research model consists of
words that towards information diffusion.
multiple literature reviews we
convey the The social network built using the
disciplines studied to determine the
various user profiles helps in answering
, such as features extraction,
information the last hypotheses. If a dense
informatio regression and graph
opinions during network is built around
n science visualization.
a data breach. technology and security focused
[13], The user profiles provide
We have groups and more links are
marketing interesting insights on the
studied established, then we can conclude
[14], interests they share.
literature from that they contribute more towards
psycholog Running a classification
the perspective information diffusion.
y [15], algorithm to segment user
of customers
among profiles into multiple
who are also 3. Research Methodology
others. groups based on the words
the victim and
Based on they used in the profiles
from the
these helps in establishing a user The research methodology for
viewpoint of
network network. Based on the validating the retrieval of
the firms in
Page
71892
message groups Bank a major
characteristics that involved in 3.1. Data Collection Canadian bank data
contributes to the sharing breach and Coca-Cola
information diffusion messages breach. We used a
Data collection
and visualizing constitutes of combination of search
is the first
various user primarily six criteria, using both
phase of our
phases as the hashtags and
methodology.
shown in the search keywords. To
To validate and
figure 1. make the data
test our
starting from collection exhaustive,
hypothesis we
collecting we have used
have collected
social media multiple keywords
the user tweets
data (step (a) in other than just “data
from social
the figure) to breach” as a single
media website
visualizing the keyword and
Twitter. The
social networks considered 7 hashtags
dataset covers a
of user groups in conjunction with
period of about
(step the search keywords.
35 days of
(f) in the The below table
tweets from
figure). The shows the list of
May 1st, 2018
steps in the search keywords and
till May 21st,
methodology of the hashtags used.
2018. The total
implementation number of
are represented Table 1.
tweets collected Search
in clockwise are 73,036 Keywords for
direction in the using various data
figure 1. All search collection
the six phases keywords and Search Keywords data breach, bre
of the hashtags. The hack
methodology breach data Hashtags #databreach #c
are explained includes #data #infosec
below from discussions #breach #hacki
sub-sections around
3.1 for (a) Facebook, 3.2. Data Pre-processing and
through Commonwealth Cleaning
3.6 for (f).
The next phase of our #breach hashtag and class
methodology includes data training notifications tagged with
cleaning and filtering hashtags #cybersecurity. These
unnecessary records. As it are identified using multiple
is with every text mining search conditions.
technique, the quality of
output depends on the 3.3. Word2Vec
quality of input used for
training the algorithm. So
The third phase of our
we have removed all the
methodology involves extracting
unnecessary special
contextually similar words. For
characters, URLs, hashtags,
this we use Word2Vec, a single
mentions, common and
layer neural network. Word2Vec
most frequent stop words
is an unsupervised machine
like the prepositions, and
learning algorithm. We have
other attribute fields like
Figure adapted the tensorflow (a Python
@RT for retweet. Some of
1. package) implementation of
the tweets we collected are
Conte Word2Vec to capture results. We
not related to the data
xt pass the input sentences as words
diagra breach. Examples of which
by tokenizing the sentences to
m include the sarcastic
words. The Word2Vec converts
messages tagged with
the high dimensional
Page
71902
representation of We have
Query word based on Semantically also
close words regression model as
words and sentences prior literature considered the below:
into low dimensional presence of
stress reaction, effect, risk, bug,
word embeddings. URLs
learn, vulnerable in the log retweet count =
The output of running message, which β0 + β1*url +
public opinion, detail, social,
the algorithm is the is usually an β2*hashtags +
prepare, reveal, key, β3*stress + β4*public
closest words of the image,
comprehend video
given word based on file or a news + β5*anxiety +
anxiety worry, sensitive, personal, β6*efficacy +
the input text. The perplex,article link,
report, compromise β7*threat + β8*hack
input text is the set of conveying the
efficacy regulate, security, account, + β9*company + ϵ
tweets related to data evidence
password, for
protect, legal
breach and the query the message
threat discovery, massive, missing, 3.5. User profile
word. If the input posted or
million, severe, upset segmentation
word is “stress”, then shared and the
hack breach, attack, stole,
based on all the discover,number
share, avoid of
tweets passed to the hashtags in Segregating users
company facebook, equifax, uber, based on their
algorithm, the closest each tweet.
arronbank, commbank, interests provides
words associated with With
cambridgeanalytica these
“stress” are being insights about the
“reaction”, “effect”, independent user groups who are
“risk”, “bug”, “learn”, 3.4. Estimation variables, we involved in sharing
and “response”. These methods have taken the the messages. For this
are the 6 words retweets count we have relied upon
closest to “stress”, as We utilized regression as the the estimation
they either appear analysis to validate dependent parameters. We
before the word or hypothesis variable for our retrieved the users
after it in the semantic (1) and (2). We regression who posted or shared
context. The window derive the model. With the messages that
used here is 3 before keywords from the retweets contain the significant
and 3 after, which the entire set of count being a keyword parameters.
makes the algorithm tweets about continuous E.g: If “stress” turns
to return data breaches, integer out to be a significant
6 closely associated by querying the variable, we parameter, we capture
words. Taking too model using have used the all the users who
many close words is the words negative posted messages
not ideal as there is a collected from binomial related to stress.
high probability of literature Using the
considering review based username from the tweet
overlapping words on prior post, we retrieved the
and it adds up to studies. These profile description from
noisy data. The below derived Twitter. We ran the
table shows all the keywords form clustering algorithm [18]
words semantically the basis for on the text description to
associated to input features of the retrieve topics that interest
word. model. We social media users who
Table 2. Search search for the tweet on data breach
Keywords for presence of messages. Table 3. shows
data collection these words in all the topic words Table 3. Interesting topics of
each and every generated from the user social media users
tweet message profiles. User Group Topic Words
and mark for its Security Awareness analytics, cyber
presence using malware, safety
binary values solutions, secur
of 1 and 0 for Technology Focus tech, infrastruc
presence and bigdata, blockc
absence of the nosql, bigdata
keywords
respectively.

Page
71912
News Related direction of the variable in are contribut
communication the overall ng positi
. Users are the model. Also, Z-
Social Work nodes of the value provided Table 4. Regression
in the table 4 results
graph and the
provide User Group Estimate Std.Err ZV
communication
Other interests valuable Intercept 0.854 0.229 3.72
link is the edge
insights about URLs 1.026 0.181 5.66
between the
the error. The Hashtags 0.745 5.602 0.13
nodes in the
higher the Stress 0.303 0.109 2.75
graph. Edges of
magnitude of Public 5.781 4.314 1.34
3.6. Social Network equal weight
Z-value there is Anxiety 0.232 0.124 1.88
Analysis (SNA) are formed
less error in the Efficacy 0.278 0.165 1.68
between user1
overall model Threat -0.469 0.121 -3.89
The last phase of our and user2, if a
due to this Hack -0.732 0.441 -1.66
methodology is to message posted
variable and it Company 0.474 0.246 1.92
create a visualization by user1 or
user2 is replied has to be
of the category of considered in Based on the above
users who share the to or shared by
user2 or user1 the overall statistical results, the
messages across model. updated model turns
social media. To gain respectively.
We color code Standard error out to be as below:
a deeper insight into is captured in log retweet count =
the information the type of user
the third β0 + β1*url +
diffusion, we build a groups in the
column with β3*stress +
social network using social network β5*anxiety +
for clear the Z Value in
the groups the users 4th which could β6*efficacy +
are clustered into. We visualization. β7*threat + β8*hack
also provide an
retrieve all the + β9*company + ϵ
estimate of
messages that contain 5. Results standard
the significant The figure 2., shows
deviation. The
characteristics in a interaction between
We exponent of
message along with various user groups.
summarize the estimate is
the information of There is wide amount
results of provided in the
user who posted, of communication
regression in last column.
replied or shared the between technology
Table 4. The The highly
message. We did not focused groups and
first column significant
build a directional security related
represents all variables of the
graph as we are experts. While there
the variables in model are
interested in finding are significant
our mode intercept,
communication contributions made
including the URLs, and
between user groups by the news channels
intercept and Threat. The
but not the and personnel from
the coefficients next significant
media, the other
are present in variable is
community users
the estimate. Stress. All
were not widely
The higher the these 4 factors
discussed their
value of
coefficient, the posts. There are users who characteristics by user groups
more is the responds to some of the provide interesting insights into
contribution of data breach messages have the information diffusion pattern.
wide range of interests that
are not specific to breaches Figure 2. Social Network
alone. Interestingly the of
users who are more various
concerned toward society user
in general responded to groups
significant amount of posts.
Thus, visualizing the posts
with significant message
6. Conclusion and
Page
71922
Limitations longitudinal 581166/privacy-
Our research provides a analysis using personal-data-report/
novel approach of finding the message
the message characteristics [2] Romanosky, S.,
posts from
that facilitates information Telang, R., & Acquisti,
various social A. (2011). Do data
diffusion. Especially, media users. breach disclosure laws
addressing data breach Restricting the reduce identity theft?.
scenarios through this posts through Journal of Policy
study helps organizations certain users by Analysis and
and governments officials time period Management, 30(2),
two folds by not only helps in 256-286.
providing insights into the performing this
kind of messages that are [3] Acquisti, A.,
analysis. We Friedman, A., & Telang,
vulnerable in the social plan to include R. (2006). Is there a cost
media platforms and the this study in to privacy breaches? An
type of user groups that our next event study. ICIS 2006
they need to address extended work Proceedings, 94.
through communication in on this data.
order to mitigate the costs We have [4] Belsis, P.,
through data breaches and considered 35 Kokolakis, S., &
helps in reestablishing lost Kiountouzis, E. (2005).
days of data
reputation. Information systems
breach security from a
messages as a knowledge management
We have utilized the test bed for perspective. Information
prior studies for creating a validating our Management &
list of query words for study, Computer Security,
capturing the semantically increasing the 13(3), 189-202.
similar words from the volume and
social media messages. [5] Choi, N., Kim, D.,
timing of these
Goo, J., & Whitmore, A.
This list could not be messages (2008). Knowing is
considered as exhaustive. toward breach doing: An empirical
Adding additional query situation and validation of the
strings to the search criteria then relationship between
could bring in more segregating the managerial information
valuable insights and could messages by security awareness and
bring in other significant industry could action. Information
features from the data. Management &
bring in
Computer Security,
Even though we have different 16(5), 484-501.
considered an automated insights. This is
manner of retrieving the another [6] Barnes, C. M., &
characteristics of breach perspective Van Dyne, L. (2009).
messages through text which we I'm tired': Differential
mining techniques and would like to effects of physical and
unsupervised machine add to follow- emotional fatigue on
learning, we did not on studies. workload management
perform the strategies. Human
Relations, 62(1), 59-92.
7. References
[7] Elhai, J. D., & Hall,
B. J. (2016). Anxiety
about internet hacking:
[1] Kedmey, D. Results from a
(2014, November community sample.
12). 9 in 10 Computers in human
Americans Feel behavior, 54, 180-185.
They've Lost
Control of Their [8] Bulgurcu, B.,
Personal Data. Cavusoglu, H., &
Time. from Benbasat, I. (2010).
http://time.com/3 Information security

Page
71932
policy (2009). How
meeting of the Information Systems,
compliance ethics can
association for 33(3), 904-933.
: an enhance
computational
empirical organizational
linguistics (pp. [24] Cavusoglu, H.,
study of privacy: lessons
384-394). Mishra, B., &
rationality- from the
Association for Raghunathan, S. (2004).
based choicepoint and
Computational The effect of internet
beliefs and TJX data
Linguistics. security breach
information breaches. Mis
announcements on
security Quarterly, 673-
[20] Wang, R. market value: Capital
awareness. 687.
Y., & Strong, D. market reactions for
MIS
M. (1996). breached firms and
quarterly, [10] Mikolov,
Beyond accuracy: internet security
34(3), 523- T., Sutskever, I.,
What data quality developers. International
548. Chen, K.,
means to data Journal of Electronic
Corrado, G. S., & consumers. Commerce, 9(1), 70-
[9] Culnan, Dean, J. (2013). Journal of 104.
M. J., & Distributed management
Williams, representations information [25] Acquisti, A.,
C. C. of words and systems, 12(4), 5- Friedman, A., & Telang,
phrases and their (2018). The moderating role 33. R. (2006). Is there a cost
compositionality. In of popular peers’ achievement to privacy breaches? An
Advances in neural goals in 5th-and 6th-graders’ [21] Chakraborty event study. ICIS 2006
information processing achievement-related , R., Lee, J., Proceedings, 94.
systems (pp. 3111- friendships: A social Bagchi-Sen, S.,
3119). network analysis. Journal of Upadhyaya, S., & [26] Stieglitz, S., &
Educational Psychology, Rao, H. R. Dang-Xuan, L. (2013).
[11] Mikolov, T., Chen, 110(2), 289. (2016). Online Emotions and
K., Corrado, G., & shopping information diffusion in
Dean, J. (2013). [16] Stieglitz, S., & Dang- intention in the social media—sentiment
Efficient estimation of Xuan, L. (2013). Emotions context of data of microblogs and
word representations in and information diffusion in breach in online sharing behavior.
vector space. ICLR social media—sentiment of retail stores: An Journal of management
Workshop. microblogs and sharing examination of information systems,
behavior. Journal of older and 29(4), 217-248.
[12] Burt, R. S., Kilduff, management information younger adults.
M., & Tasselli, S. systems, 29(4), 217-248. Decision Support [27] Cavusoglu, H.,
(2013). Social network Systems, 83, 47- Mishra, B., &
analysis: Foundations [17] Lee, J., Agrawal, M., & 56. Raghunathan, S. (2004).
and frontiers on Rao, H. R. (2015). Message The effect of internet
advantage. Annual diffusion through social [22] Patel, A., security breach
review of psychology, network service: The case of Qassim, Q., & announcements on
64, 527-547. rumor and non-rumor related Wills, C. (2010). market value: Capital
tweets during Boston A survey of market reactions for
[13] Otte, E., & bombing 2013. Information intrusion breached firms and
Rousseau, R. (2002). Systems Frontiers, 17(5), 997- detection and internet security
Social network analysis: 1005. prevention developers. International
a powerful strategy, also systems. Journal of Electronic
for the information [18] Blei, D. M., Ng, A. Y., & Information Commerce, 9(1), 70-
sciences. Journal of Jordan, M. I. (2003). Latent Management & 104.
information Science, dirichlet allocation. Journal Computer
28(6), 441-453. of machine Learning Security, 18(4),
research, 3(Jan), 993-1022. 277-290.
[14] Ansari, A., Stahl,
F., Heitmann, M., & [19] Turian, J., Ratinov, L., & [23] Choi, B. C.,
Bremer, L. (2018). Bengio, Y. (2010, July). Word Kim, S. S., &
Building a social representations: a simple and Jiang, Z. (2016).
network for success. general method for semi- Influence of
Journal of Marketing supervised learning. In firm’s recovery
Research. Proceedings of the 48th endeavors upon
annual privacy breach on
[15] Laninga-Wijnen, online customer
L., Ryan, A. M., behavior. Journal
Harakeh, Z., Shin, H., & of Management
Vollebergh, W. A.
Page
71942

You might also like