Content uploaded by Ponnurangam Kumaraguru
Author content
All content in this area was uploaded by Ponnurangam Kumaraguru on May 04, 2021
Content may be subject to copyright.
Methods for User Profiling Across Social Networks
Rishabh Kaushal1,2, Vasundhara Ghose1, and Ponnurangam Kumaraguru2
1Department of Information Technology, Indira Gandhi Delhi Technical University for Women, Delhi, India
2Precog Research Group, Indraprashtha Institute of Information Technology, Delhi, India
Abstract—Users have their accounts on multiple Online Social
Networks (OSNs) to access a variety of content and connect
to their friends. Consequently, user behaviors get distributed
across many OSNs. Collection of comprehensive user information
referred to as user profiling; an essential first step is to link
user accounts (identities) belonging to the same individual across
OSNs. To this end, we provide a detailed methodology of five
methods useful for user profiling, which we refer to as Advanced
Search Operator (ASO), Social Aggregator (SA), Cross-Platform
Sharing (CPS), Self-Disclosure (SD) and Friend Finding Feature
(FFF). Taken together, we collect linked identities of 208,120
individuals distributed across 43 different OSNs. We compare
these methods quantitatively based on social network coverage
and the number of linked identities obtained per-individual.
And also perform a qualitative assessment of linked user data,
thus obtained by these methods, on the criteria of completeness,
validity, consistency, accuracy, and timeliness.
Index Terms—User Profiling, Social Media Analysis, Online
Social Networks.
I. INTRODUCTION
The popularity of Online Social Networks (OSNs) is in-
creasing by the day with more and more people joining
multiple OSNs to share information about themselves, connect
to other users, and receive updates from them. Each OSN
offers a unique service or ecosystem which attracts users to
join more than just one of them. Facebook for personal friends,
LinkedIn for the professional network, YouTube for viewing &
sharing videos, and Twitter to get quick updates are the best
options. The average number of social media accounts per
online user has risen from 4.3 to 7.6 during 2013 to 20171.
To collect user information in a comprehensive manner, an
essential first step is to gather user accounts (identities) of the
same individual across multiple OSNs, which we refer to as
linked identities. And the systematic approach to performing a
large scale collection of user behaviors across OSNs is referred
to as user profiling [1].
User profiling has many advantages and applications. Users
tend to provide incomplete information on a single social
network, either with purpose or otherwise. Knowing the same
user’s identity on other social networks would help in the com-
prehensive profiling of the user in terms of user’s profile, user’s
content, user’s behavior, user’s preferences, and user’s friends.
In the advertising world, it enables targetted advertisement
[2] and improved recommendations. Researchers have studied
most of the problems in the domain of social networks like
1https://www.statista.com/statistics/788084/number-of-social-media-
accounts/
information propagation, link prediction, algorithmic biases,
discrimination studies, and community detection in the realm
of a single social network, which we can study across multiple
social networks. In social media crimes and cybersecurity
problems like cyberbullying, fake accounts, and spamming, we
are often looking for user footprints within the same social
network in which incident occurred. If the user’s identities
on other social networks are known, then it is only going to
help in the investigation [3]. From a user’s privacy standpoint,
individuals can be shown their comprehensive profiles and
likelihood of linkage of their identities and nudged to control
their online behavior so that their digital footprint decrease [4].
Lastly, there is no agreed benchmark dataset in the problem
domain of identity resolution. So, large scale data collection of
linked identities would help researchers compare and evaluate
their proposed solutions.
Given the significance of user profiling, a lot of emphases
has been given in the research community to solve the first
step in user profiling which involves linking user identities
belonging to the same person, referred to as identity resolution
(or identity linkage).2A data-driven approach to solve the
identity resolution problem has two key steps. Firstly, we
collect a large number of user identity pairs belonging to
linked identities and non-linked identities. Secondly, we con-
struct a machine learning-based model over the user behavioral
features extracted from user identities. In this paper, we focus
our attention on the first step, which involves the collection
of linked identities across OSNs. Figure 1 depicts the before
and after stages involved in linked identity collection for user
profiling. In this paper, we explain five methods to obtain
linked identities namely Advanced Search Operator (ASO),
Social Aggregator (SA), Cross-Platform Sharing (CPS), Self-
Disclosure (SD) and Friend Finding Feature (FFF). Taking all
these methods together, we collect linked identities of 208,120
individuals across 43 different OSNs, which is by far the
most comprehensive coverage, towards user profiling, refer
at http://precog.iiitd.edu.in/resources.html for dataset details.
Subsequently, we present a detailed quantitative and qualitative
assessment of these methods. For quantitative assessment,
we evaluate the number of social networks covered by a
method and number of linked identities obtained per-individual
across OSNs. For qualitative assessment, we leverage standard
2This problem is known in literature by multiple names such as Social
Identity linkage [5], User identity linkage [6], user Identity Resolution Social
Network Reconciliation [7] , User Account Linkage Inference [8], Profile
Linkage [9], Anchor Link prediction [10] and Detecting me edges [11].
Fig. 1. Visual depiction of progressive stages in which linked identities are collected starting from no linked identities and gradually progressing to collect
as many of them as possible by applying methods for user profiling.
parameters from ISO 9000:20153namely data completeness,
consistency, accuracy, validity, availability and timeliness.
Collecting user data from online social networks have
always been a challenge and given that the data is related
to users, there are privacy issues as well. Application Pro-
grammer’s Interfaces (APIs) offered by OSNs have been
dwindled their capabilities over the years owing to data privacy
concerns. The recent data breach4involving Facebook and
Cambridge Analytica would have an adverse implication on
data collection by academics for research purposes.5With
all this happening, users are becoming even more privacy
aware which would dissuade them from mentioning all the
details in their accounts, resulting in missing values when
data is collected. To make things worse, there are social
network platforms like Twitter which allow users to change
their account handles, thereby, complicating the data collection
process.
Regardless, this is the first work to the best of our knowl-
edge which focuses exclusively upon the methods for collect-
ing linked identities, which is the the de-facto first step for
user profiling. Key contributions of our work are:-
•Detailed description of data collection methods to retrieve
linked identities, thereby facilitating user profiling.
•Comprehensive evaluation of data collection methods
both qualitatively and quantitatively.
•Creation of a comprehensive dataset that can be used as
benchmark dataset for identity resolution research.
II. RELATED WORK
One of the earliest works is that of Perito et al. [12],
who conducted various data collection methods to obtain the
usernames belonging to the same individual across many sites.
Prominent sites used were Google profiles and eBay to collect
3.5 million usernames and 6.5 million usernames, respectively.
For ground truth, they relied on Google profile users who
3International Standards Organization: https://www.iso.org/standard/45481.html
4https://www.theguardian.com/news/2018/mar/17/cambridge-analytica-
facebook-influence-us-election
5https://scroll.in/article/872770/cambridge-analytica-scandal-could-hurt-
legitimate-researchers-using-facebook-data
have listed their usernames on other sites. Malhotra et al. [13]
used Google API to retrieve a list of user accounts declared
by users on their Google profile. They collected ground truth
from Twitter, YouTube, and Flickr; however, there were many
missing fields. Twitter and LinkedIn were other platforms to
obtain 29,129 pairs of user accounts. Zafarani et al. [14],
in their work, collected usernames of the same real world
user across 32 different sites. Three primary sources to obtain
these matching usernames were namely social network sites
like Facebook or Google+ where online users mention their
usernames on other websites, blogs, or blog advertisement
portals. Oana Goga et al. [15] collected identical usernames
on three of the social network being considered Twitter,
Yelp and Flickr using ‘Friend Finder’ mechanism present
on these sites. Further, in their work [16], they focused
on Twitter, Facebook, Google+, Flickr, and Myspace. User
features from these social networks were obtained using their
APIs. However, for ground truth, 3 million Google+ profiles
were randomly crawled to exploit the fact that user list down
their social network accounts on Google+ profile pages. Iofciu
et al. [17] used Social Graph API to crawl 421,188 public
profiles of users while considering Flickr and Delicious and
StumbleUpon (FDS dataset). Kong et al. [10] considered
Twitter and Foursquare as the two real-world networks for
their investigation. For ground truth, they used users who had
mentioned in Twitter ID’s in their Foursquare account pages.
Xin Mu et al. [6] considered Chinese social networks Weibo,
Renren, 36.cn, and Zhaopin for studying identity resolution.
For ground truth, they annotated 2,186 pairs of user accounts
across these social network pairs. Man et al. [18] took their
first dataset from crawling of Facebook users, after deleting
the users who have less than five friends, there remains 40,710
users with 766, 519 connections. The second data set was co-
author network formed from papers published in conferences
in the domain of Data Mining and Artificial Intelligence.
Peled et al. [19] collected data using web crawling from two
social networks, Xing and Facebook. They manually obtained
seed profiles for crawling pairs of user profiles, one from
each network, that belong to the same individual. A tool was
Fig. 2. Generic Framework for User Profiling.
developed to extract user data like gender, name, education,
professional experience, friend list, etc. Labitzke et al. [20]
collected 110,000 Facebook profiles, more than 43,000 profiles
of StudiVZ, more than 25,000 MySpace profiles and more than
10,000 profiles of Xing. Liu et al. [21] collected the dataset by
performing a survey based on 153 respondents and an analysis
of 75,472 users on About.me website. For the collection of
ground truth, they hired one human annotator who considered
user profile content pages with more details related to users
and their posts to mark true positive identities. Riederer et
al. [22] collected datasets from location-based data extrac-
tion from OSN. Most interestingly, they included Call Data
Records (CDRs) as well for making use of cell phone location
tracking. OSNs considered were Foursquare-Twitter-Instagram
cellphone-credit card records. They collected ground truth
from the publicly available dataset and cell-bank. Zhang et
al. [9] used a synthetic network based on Renren, Facebook,
Sima, WeChat, and Twitter. They collected ground truth by
using page crawling and open API wherever possible as per
the platform is chosen.
Our work is different from prior works in the sense that we
focus on methods to collect linked identities to enable the user
profiling and present comparative assessment of these methods
on qualitative and quantitative parameters.
III. METHODOLOGY
A generic framework for user profiling (Figure 2) comprises
three steps, namely data collection, data integration, and data
extraction & indexing. The first step is data collection, in
which we identify a source of data followed by a selection
of data collection methods. We follow it by data integration
in which we store user identities collected from all methods
at a single data store point, which we refer to as Linked
Identity Data Store (LIDS). Finally, data extraction and in-
dexing involve collecting the three components of user identity
namely, profile, content, and network. Next, we describe each
data collection method in detail. Next, we present a detailed
methodology adopted to perform data collection using five
methods, which is the focus of this paper.
A. Advanced Search Operator (ASO)
Search engines typically provide advanced search oper-
ators using which users use to obtain more detailed and
specific information. In this work, we leverage Google’s
advanced operator search, also referred to as google hack-
ing or google dorking (Figure 3). For instance, the search
query intext:facebook.com,twitter.com filetype:xlsx would
locate all web documents that have facebook.com or
twitter.com written as text anywhere in the record with
the additional constraint that these documents must be of
xlsx file type. As per Figure 3, we first run a script which
searches using pre-configured search queries on a specific
search engine. Downloaded files are filtered and subsequently
read through automated scripts which store linked identities in
LIDS.
Fig. 3. Pipeline for Advanced Search Operator (ASO) method.
B. Social Aggregator (SA)
There are several social aggregating websites on which users
create an account and provide details of their multiple OSN
accounts. One such site that we investigate is about.me6which
is a website that offers its users with a platform to mention
numerous user identities, external websites, and well-known
social networking websites such as Facebook, Flickr, Google+,
Pinterest, LinkedIn, Twitter, Tumblr, and YouTube. Users put
their one-page descriptions giving details of their social media
profiles along with their background image and abbreviated
biography. Initially, when we started data collection using
this method, about.me provided an option to search user
profile using the topic-based search (referred to as discovery
feature). Given an interest-topic as input, it would return all
the user profiles having that interest. After one month of data
6About.me: https://about.me/
collection, in March 2018, we found that this discovery feature
of about.me got discontinued. Subsequently, on exploring
other options, we found a public dataset7containing about.me
profiles which we use in this work.
Fig. 4. Pipeline for Social Aggregator (SA) Method.
Besides, we leverage the previous ASO method using in-
terests as intext and site as about.me to obtain more user
profiles. Figure 4 explains the three data sources employed
in this method.
C. Cross-Platform Sharing (CPS)
Many OSNs provide an option to share content across other
(target) OSNs which we refer as cross-platform sharing (CPS).
As depicted in Figure 5, a user makes a post on the source
network (say Zomato, Facebook or Instagram) and then subse-
quently shares the same post on the target network (Twitter).
Such shared content on the target network appears with a
specific pattern. In our work, when we took Twitter as the
target network and Instagram as the source network, then the
pattern that appears on the shared post is \instagram.com\p\.
Using the API provided by the source network, we search
for posts that contain such patterns. Besides this pattern,
we also specifically check for the source field present in
the Tweet JSON object and make sure that it has the name
of the source network (in our case, Instagram). This check
ensures that we filter out those scenarios in which a user might
have copy-pasted the URL pattern because of such situations
are not guaranteed to link to the same individual across the
two networks. We parse the collected posts from the target
network, identify the URL and expand the URL to reach the
desired content on the source social network. On reaching the
source social network, we either use source social network API
or scrap the post page to obtain the tagged user (mentioned
user) in the post on the source social network. In this way, we
obtain linked identity pair between source and target social
network.
7http://scholarbank.nus.edu.sg/bitstream/10635/137403/2/about me.sql
Fig. 5. Pipeline for Cross Platform Sharing (CPS) method.
D. Self-Disclosure (SD)
Whenever a user signs up on OSN, there is an option to
provide a user description. At times, users provide details of
their identities on other OSNs, which we refer to as self-
disclosure. More specifically, we focus on the user’s bio field
in the Twitter network (Figure 6). We first use Twiangulate
web tool8to collect all those twitter profiles which have at
least one social network mentioned in their bio-field. Then,
we observe various patterns in the bio-field on Twitter because
a user can specify other OSN details in multiple ways. For
instance, a user can mention TV Host and Media Trainer
- Instagram: @NeshanTVxyz Snapchat: @Neshaxyz while
another user can use acronyms like TV Host and Media Trainer
- IG: @NeshanTVxyz SP: @Neshaxyz FB: nashbin123. To
address these variations, we tokenize all text and check for
the occurrence of URL which could lead to other OSNs.
E. Friend Finder Feature (FFF)
Whenever a user joins a new OSN, we sign up using our
unique identifier, say email or phone number. This information
is used by OSN to find our friends in our email contacts or
phone contacts. Using this information, OSN offers a friend
finder option to help connect to those friends who already
have an account in OSN. Figure 7 depicts the entire sequence
of steps that we followed in this method. In the first step, we
use a deep web search engine like Duckduckgo 9for retrieving
emails present over the web. Next, we create an email account
and add the extracted in its contact list. Then we sign-up in a
social network to exploit friend finder feature using the created
account. We use string matching on display name of users to
find identity belonging to the same user.
IV. RES ULT S AN D EVALUATI ON
In this section, we compare five methods, as stated before,
by performing a quantitative and qualitative assessment of
the linked identities obtained by them. Table I depicts total
identities collected by prior works along with the OSNs
covered by them. Few of them have a higher number of
8Twiangulate: http://twiangulate.com/search/
9Deep Web: www.duckduckgo.com
Fig. 6. Pipeline for Self-Disclosure (SD) Method.
Fig. 7. Pipeline for Friend-Finder Feature method.
TABLE I
IDENTITIES COLLECTED IN PREVIOUS WORKS.
OSNs Covered with Reference Identity collected
Twitter, Foursquare and yelp [15] 17,276
Twitter, Flicker, Facebook, Google+,
Myspace, Yelp [16] 655,079
Social Graph API [17] 421,188
Facebook Twitter [10] 500
StudiVZ, Facebook, Myspace and Xing [7] 89,000
Facebook and Myspace [23] 5,296
Twitter, Flicker and Linkedin and 12 more.. [21] 75,472
Weibo, Renren, 36.cn and Zhoopin [6] 25,647
Facebook and Xing [24] 158
Twitter, Flicker [19] 27,000
32 Social Network sites [14] 100,179
Foursquare, Twitter and Instagram [22] 2,579
Twitter, YouTube and Flicker [13] 41,336
Twitter and BlogCatlog [25] 3,000
Facebook and MAG [18] 1,154
identities; however, their coverage in terms of the number
of OSNs reached is less than our dataset. Additionally,
datasets of prior works as mentioned in Table I are not
publicly available while we release our dataset, more details at
http://precog.iiitd.edu.in/resources.html . Table II summarizes
the number of linked identities collected using each of the
data collection methods. Among all the five methods, Cross-
Platform Sharing (CPS) method yielded the maximum number
TABLE II
RES ULTS O F DATA COLL EC TIO N MET HO DS IM PL EME NT ED IN T HI S WOR K.
DATA COLLECTION IN EACH OF THEM IS CONTINUING AND NUMBERS ARE
IN CRE AS ING B Y THE D AY.
Data Collection Method Linked Identities
Collected So Far
Advanced Search Operator (ASO) 9,695
Social Aggregator (SA) 53,692
Cross-Platform Sharing (CPS) 104,233
Self Disclosure (SD) 40,000
Friend Finder Feature (FFF) 500
Total Linked Identities 208,120
of linked identities (104,233) keeping Twitter as the target
network and Zomato, Facebook, and Instagram being the
source network from where the post was shared on to Twitter.
Social Aggregator (SA) method using about.me gave 53,692
linked identities taking into account all three approaches
followed in it, namely discovery feature, which contributed
15,973, a standard dataset that added 15,620 and search engine
based which yielded 22,099. Self Disclosure (SD) method,
which extracted identities by parsing bio field of Twitter gave
40,000 linked identities. We collected 9,695 identities using
Advanced Search Operator (ASO) queries on google. Lastly,
Friend-Finder Feature (FFF) gave 500 linked identities.
Fig. 8. Distribution of coverage of OSNs on which linked identities got
collected using Advanced Search Operator (ASO) and Self Disclosure (SD)
methods. Values on Y-axis are on log-scale to the base 10.
A. Quantitative Evaluation
For quantitative evaluation, we evaluate data collection
methods based on two metrics explained below.
1) Social Network Coverage: The data collection method
is intended to collect linked identities across as many OSNs
as possible. Social network coverage refers to the number of
OSNs on which the given data collection method was able
to collect linked identities. As depicted in Figure 8 Advanced
Search Operator (ASO) method covers nine social networks
Facebook, Twitter, Youtube, Linkedin, Google+, Pinterest,
Instagram, Soundcloud, and Twiplomacy, during coverage of
Self Disclosure (SD) method is across four social networks
Twitter, Facebook, Instagram, and Snapchat (others comprises
of LinkedIn and their blog/ websites). Further, in terms of
OSNs coverage, Social Aggregator (SA) method performs
the best. As depicted in Figure 10, a total of 43 OSNs
got covered using this method. Among the three approaches
employed in the SA method, the one that leverages search
engine (duckduckgo) is giving the best results.
2) Per-user linked identity count: Number of linked iden-
tities found for a given user is referred to as per-user linked
identity count. Figure 9 depicts the number of linked identities
per user obtained using Advanced Search Operator (ASO)
and Self Disclosure (SD) method. For per-user linked identity
count less than 4, SD performs better, but subsequently ASO
performs well. Also as expected, with the increase in per-user
linked identity count, the number of such users, decrease. Fig
11 shows the per-user identity count distribution for Social
Aggregator (SA) method. Discover feature and public dataset
approaches give better results during the ongoing approach of
the search engine is providing comparable results with discov-
ery feature when per-user identity count increases beyond 5.
3) Results of Cross-Platform Sharing (CPS) Method:
Quantitatively, Cross Platform Sharing (CPS) method is giving
the best results, in terms of number of linked identities
Fig. 9. Distribution of per-user identity count using all methods except social
aggregator (SA) method.
TABLE III
RES ULTS O F CROSS PL ATFO RM SHARING (CPS) METHOD IN WHICH WE
DE PIC T DI STR IBU TI ON OF C ROS S PL ATFOR M SH ARE D PO STS F ROM T HRE E
SO URC E NE TWO RKS N AME LY ZOMATO , FACE BOO K AND IN STAG RAM O N
TWITTER.
Source Network Linked Identities
Zomato 6,000
Facebook 40,201
Instagram 58,032
Total Linked Identities 104,233
obtained. In the CPS method, we have used Twitter as the
target network on which posts from other source networks
namely Zomato, Facebook and Instagram have been shared,
Table III gives the distribution of the same. Cross platforms
sharing from Instagram to Twitter got the most of the linked
identities.
B. Qualitative Evaluation
We leverage metrics from ISO 9000:201510 Standard for
quality assessment, namely data completeness, consistency,
accuracy, validity, availability, and timeliness.
•Completeness: Completeness, in our context, can be
defined as the ratio of collected linked identities of a
user to the actual linked identities across all OSNs for the
same user. From information retrieval perspective, this is
similar to recall. Ideally, the methods should contain all
linked identities but in practice, it is not possible, refer
Table IV for explanations.
•Validity: Validity in the context of linked identity col-
lected would mean whether the collected identity pair
indeed belong to the same user in the real world. We
are expected to get valid linked identities as long as
the users keep their identity lists and profile descriptions
correctly updated in methods namely Social Aggregators
(SA) and Self Disclosure (SD), respectively. In the case
of the Advanced Search Operator (ASO) method, if the
10International Standards Organization: https://www.iso.org/standard/45481.html
Fig. 10. Distribution of social network covered using Social Aggregator (SA) method for collection of linked identities. This method by far is the best in
terms of OSN coverage with total 43 OSNs covered.
Fig. 11. Distribution of per-user linked identity count using Social Aggregator
(SA) method for collection of linked identities. It may be noted that 24 users
identity count more than 20 have not been plotted in this graph to keep
visualization comprehensible.
TABLE IV
COMPLETENESS ANALYS IS O F DATA COLLECTION METHO DS.
Method Remarks on Completeness
ASO Depends on number of social identities submitted by
user to any server whose data is indexed by search
engines
SA Depends on number of social identities displayed by
user on social aggregator sites
CPS Depends on cross sharing activity of user
SD Depends on amount of URLs mentioning identities
on other OSNs revealed by user in his/her account
description
FFF Depends on availability of friend-finder feature on
OSN and friends having account on that OSN
indexed file is quite outdated, then linked identities could
be stale.
•Consistency: While each of the data collection methods
would execute consistently, however, due to the dynamics
of the OSNs, the results for each run could vary. Some
OSNs provide greater re-configurability in user profiles,
for instance, username can be changed in Twitter and
Instagram. If a given data collection is relying upon
username, then results would vary over time.
•Accuracy: All the methods rely upon user-contributed
information. In the case of Advanced Search Operator
(ASO), it is the data entered into servers which are
indexed by search engines whereas in Social Aggrega-
tor (SA) and Self Disclosure (SD) methods is directly
dependent on the information provided by the user. As
long as user-supplied information is accurate, the data
collection methods are guaranteed to return true positive
linked identities.
•Timeliness: In the context of our problem, timeliness
would mean whether linked identities for a given input
user can be provided by the data collection method
whenever requested. Out of the five methods, methods
namely Self Disclosure (SD), Cross Platform Sharing
(CPS) and Friend Finder Feature (FFF) method could be
employed in such a situation.
V. CONCLUSION
Users are joining multiple on-line social networks for dif-
ferent purposes. An essential first step in the social profiling of
users is to link their identities across OSNs. Besides profiling,
other important applications include recommendations and
link prediction. In this work, we explained five data collection
methods and compared them both qualitatively and quantita-
tively. Based on our experience of collecting linked identities
across multiple social networks, we list down few suggestions
for prospective researchers. Social Aggregator (SA) method is
useful in the scenario when we want to study user behavior
across a large number of OSNs. Self Disclosure (SD) method
would yield a good coverage of OSNs but in a limited manner.
On the contrary, if one has to target only a specific pair of
OSN, then Cross Platform Sharing (CPS) method would be a
good option. Advanced Search Operator (ASO) method would
be useful if only popular social networks (like Facebook,
LinkedIn, Twitter, etc) are to be targeted. Friend Finder Feature
(FFF) is applicable only when a large pool of emails are
available. FFF would also be useful in the scenario when one
has to investigate an unexplored social network.
There are a few limitations to our work. For Social Aggre-
gator (SA) method, we have investigated about.me, it would
be interesting to extend it over other platforms like Google+.
Similarly in Advanced Search Operator (ASO) method, we
may go beyond google search engine and explore another
search engines like bing, duckduckgo, etc. In Cross Platform
Sharing (CSP) method, we have taken Twitter as the target
social network, which can be extended to include other OSNs
as well. Similarly, only Twitter’s bio field is being parsed in
Self Disclosure (SD) method.
Finally, for ethical reasons, all data collection methods in
this paper operate on public data only and rely upon the fact
that the user has shared this data explicitly at some point in
time. However, users may not be aware of the implications
of public availability of their data. For users who are privacy
concerned and would not want their identities to be linked, it
is highly recommended that they should not cross-post content
across OSNs, not provide details of other OSNs on their social
media profile pages, use a specific email (not known to their
friends) for registering at OSN and not register on websites
whose robots.txt allows crawling. However, regardless, this
work is a step towards a tool [4] that can help users understand
and control the amount of their data that is available on OSNs
so that they could safeguard themselves from online profiling.
REFERENCES
[1] B. Krulwich, “Lifestyle finder: Intelligent user profiling using large-scale
demographic data,” AI magazine, vol. 18, no. 2, pp. 37–37, 1997.
[2] W.-S. Yang, J.-B. Dia, H.-C. Cheng, and H.-T. Lin, “Mining social
networks for targeted advertising,” in Proceedings of the 39th Annual
Hawaii International Conference on System Sciences (HICSS’06), vol. 6.
IEEE, 2006, pp. 137a–137a.
[3] M. Huber, M. Mulazzani, M. Leithner, S. Schrittwieser, G. Wondracek,
and E. Weippl, “Social snapshots: Digital forensics for online social
networks,” in Proceedings of the 27th annual computer security appli-
cations conference. ACM, 2011, pp. 113–122.
[4] R. Kaushal, S. Chandok, P. Jain, P. Dewan, N. Gupta, and P. Ku-
maraguru, “Nudging nemo: Helping users control linkability across
social networks,” in International Conference on Social Informatics.
Springer, 2017, pp. 477–490.
[5] S. Liu, S. Wang, F. Zhu, J. Zhang, and R. Krishnan, “Hydra: Large-
scale social identity linkage via heterogeneous behavior modeling,” in
Proceedings of the 2014 ACM SIGMOD international conference on
Management of data. ACM, 2014, pp. 51–62.
[6] X. Mu, F. Zhu, E.-P. Lim, J. Xiao, J. Wang, and Z.-H. Zhou, “User
identity linkage by latent user space modelling,” 2016.
[7] N. Korula and S. Lattanzi, “An efficient reconciliation algorithm for
social networks,” Proceedings of the VLDB Endowment, vol. 7, no. 5,
pp. 377–388, 2014.
[8] Y. Shen and H. Jin, “Controllable information sharing for user accounts
linkage across multiple online social networks,” in Proceedings of the
23rd ACM International Conference on Conference on Information and
Knowledge Management. ACM, 2014, pp. 381–390.
[9] H. Zhang, M.-Y. Kan, Y. Liu, and S. Ma, “Online social network profile
linkage,” in Asia Information Retrieval Symposium. Springer, 2014, pp.
197–208.
[10] X. Kong, J. Zhang, and P. S. Yu, “Inferring anchor links across multiple
heterogeneous social networks,” in Proceedings of the 22nd ACM
international conference on Information & Knowledge Management.
ACM, 2013, pp. 179–188.
[11] F. Buccafurri, G. Lax, A. Nocera, and D. Ursino, “Discovering links
among social networks,” in Joint European Conference on Machine
Learning and Knowledge Discovery in Databases. Springer, 2012,
pp. 467–482.
[12] D. Perito, C. Castelluccia, M. A. Kaafar, and P. Manils, “How unique
and traceable are usernames?” in International Symposium on Privacy
Enhancing Technologies Symposium. Springer, 2011, pp. 1–17.
[13] A. Malhotra, L. Totti, W. Meira Jr, P. Kumaraguru, and V. Almeida,
“Studying user footprints in different online social networks,” in Ad-
vances in Social Networks Analysis and Mining (ASONAM), 2012
IEEE/ACM International Conference on. IEEE, 2012, pp. 1065–1070.
[14] R. Zafarani and H. Liu, “Connecting users across social media sites:
a behavioral-modeling approach,” in Proceedings of the 19th ACM
SIGKDD international conference on Knowledge discovery and data
mining. ACM, 2013, pp. 41–49.
[15] O. Goga, H. Lei, S. H. K. Parthasarathi, G. Friedland, R. Sommer, and
R. Teixeira, “Exploiting innocuous activity for correlating users across
sites,” in Proceedings of the 22nd international conference on World
Wide Web. ACM, 2013, pp. 447–458.
[16] O. Goga, D. Perito, H. Lei, R. Teixeira, and R. Sommer, “Large-scale
correlation of accounts across social networks,” University of California
at Berkeley, Berkeley, California, Tech. Rep. TR-13-002, 2013.
[17] T. Iofciu, P. Fankhauser, F. Abel, and K. Bischoff, “Identifying users
across social tagging systems.” in ICWSM, 2011.
[18] T. Man, H. Shen, S. Liu, X. Jin, and X. Cheng, “Predict anchor links
across social networks via an embedding approach.” in IJCAI, 2016, pp.
1823–1829.
[19] O. Peled, M. Fire, L. Rokach, and Y. Elovici, “Entity matching in online
social networks,” in Social Computing (SocialCom), 2013 International
Conference on. IEEE, 2013, pp. 339–344.
[20] S. Labitzke, I. Taranu, and H. Hartenstein, “What your friends tell others
about you: Low cost linkability of social network profiles,” in Proc. 5th
International ACM Workshop on Social Network Mining and Analysis,
San Diego, CA, USA, 2011, pp. 1065–1070.
[21] J. Liu, F. Zhang, X. Song, Y.-I. Song, C.-Y. Lin, and H.-W. Hon, “What’s
in a name?: an unsupervised approach to link users across communities,”
in Proceedings of the sixth ACM international conference on Web search
and data mining. ACM, 2013, pp. 495–504.
[22] C. Riederer, Y. Kim, A. Chaintreau, N. Korula, and S. Lattanzi, “Linking
users across domains with location data: Theory and validation,” in
Proceedings of the 25th International Conference on World Wide Web.
International World Wide Web Conferences Steering Committee, 2016,
pp. 707–719.
[23] M. Motoyama and G. Varghese, “I seek you: searching and matching
individuals in social networks,” in Proceedings of the eleventh interna-
tional workshop on Web information and data management. ACM,
2009, pp. 67–75.
[24] A. Narayanan and V. Shmatikov, “De-anonymizing social networks,” in
2009 30th IEEE Symposium on Security and Privacy, May 2009, pp.
173–187.
[25] Y. Nie, Y. Jia, S. Li, X. Zhu, A. Li, and B. Zhou, “Identifying users
across social networks based on dynamic core interests,” Neurocomput-
ing, vol. 210, pp. 107–115, 2016.
