=Paper=
{{Paper
|id=Vol-1939/paper2
|storemode=property
|title=Concealing Interests of Passive Users in Social Media
|pdfUrl=https://ceur-ws.org/Vol-1939/paper2.pdf
|volume=Vol-1939
|authors=Yaroslav Nechaev,Francesco Corcoglioniti,Claudio Giuliano
|dblpUrl=https://dblp.org/rec/conf/semweb/NechaevCG17
}}
==Concealing Interests of Passive Users in Social Media==
https://ceur-ws.org/Vol-1939/paper2.pdf
Concealing Interests of Passive Users in Social Media
Yaroslav Nechaev1,2 , Francesco Corcoglioniti1 , and Claudio Giuliano1
1 Fondazione Bruno Kessler, Via Sommarive 18, 38123 Trento, Italy
2 University of Trento, Via Sommarive 14, 38123 Trento, Italy
{nechaev,corcoglio,giuliano}@fbk.eu
Abstract. User profiling has existed in the social media since their inception and
has supported most of their business model. Even if users do not actively share the
information about themselves on the social media (so-called passive users), they
can still be profiled based on their location and who they follow. In this paper,
we present a system that leverages the linking of followed (popular) Twitter users
to DBpedia, and the information therein contained, to help users concealing their
digital footprint. Specifically, our approach helps a passive Twitter user to stay
private by proposing a list of additional profiles to follow that would confuse the
social media’s inference pipeline and prevent it from inferring useful information
about that passive user and his interests.
Keywords: Social Media, User Profiling, Linked Open Data, Machine Learning
1 Introduction
Currently, an enormous amount of people use social media every day: just recently, in
July 2017, Facebook has hit two billion monthly users. Every action of those people
is being recorded, analyzed and possibly sold to third parties in one form or another.
Additionally, this data is used to acquire a digital footprint of users: what they like or
do not like, their level of education, gender, race and much more.
Knowing that, people have learned to be careful about what they post, like or share
on social media. Some go even further — they just follow a number of profiles they
like and never actually generate any content that could be gathered or analyzed. In the
literature, such users are called “passive users”. A number of recent studies [3,7,20,26]
have demonstrated that, despite their best effort, passive users can still be profiled based
on location and the profiles they read. By adopting the paradigm “we are what we read”,
social media can infer digital footprint of passive users almost as good as for active ones,
with the result that protecting the privacy of passive users remains an open issue.
Users can, of course, choose to stop using the social media altogether in an attempt
to preserve their privacy. However, in the modern world, it is becoming increasingly
hard for an individual to abandon the benefits such services provide just for the sake of
privacy. This situation is not unique to the social media: the same happens with recent
machine learning-based consumer products such as voice assistants, translation services
and even self-driving cars. By using those services, people agree to provide data that is
used to train the system or perform inference of user profile attributes.
On the other hand, the privacy of the user is not something that has to be given up in
favor of new exciting technologies and services. Companies can choose to protect user’s
Fig. 1. The proposed concealing approach
privacy without degrading the user experience by implementing various techniques,
such as differential privacy [6] or by shifting the computation on the user-controlled
device [14]. Despite the availability of such techniques, modern social media are re-
luctant to implement them: their business depends on their ability to learn as much as
possible about the target user and exploit this information the best they can to show
advertisement and sell auxiliary services.
In this paper, we present a system that helps Twitter passive users to conceal their
digital footprint by leveraging the alignments from Twitter profiles to DBpedia entities
provided by SocialLink [17] and the categorical information available about those en-
tities. Specifically, our main contribution are two concealing approaches (Greedy and
Joint) that help passive users to stay private by proposing additional Twitter profiles to
follow (followees) that would turn the user interests distribution inferred from followees
close to the uniform one. This has the effect of confounding the social media’s infer-
ence pipeline, preventing it from inferring useful information about the real interests of
the target user. The original list of followees could then be stored on a user-controlled
device (using a custom application or a browser plugin) allowing to recreate the origi-
nal timeline. Our system proposes as few followees as possible to circumvent possible
social media-induced limitations, reduce network load and cluttering of the user time-
line. Our proposal is highlighted in Figure 1. This task of concealing user interests is
inspired by the obfuscation examples provided in Brunton and Nissenbaum’s book [5],
specifically, the “Bayesian Flooding” idea by Kevin Ludlow [13].
To evaluate our approaches we build a user’s interests inference pipeline mirroring
state-of-the-art approaches [3,20] that use the list of followees to infer passive users’ in-
terests. This pipeline is based on the idea, initially introduced by Piao and Breslin [20],
that the followee list could be mapped to the distribution of interest categories for the
user by linking followees to DBpedia/Wikipedia and then exploit the categorical infor-
mation therein contained, averaging it along followees to derive an interests distribution
for the target passive user. Our pipeline reiterates on this idea but uses the high-quality
alignments of SocialLink [17] instead of the simple heuristics used in state-of-the-art
approaches to map Twitter profiles to DBpedia/Wikipedia resources.
We evaluate our Greedy and Joint approaches and a Random baseline against this
inference pipeline. We showcase a number of results, demonstrating that the Joint ap-
proach is able to achieve almost a perfect uniform distribution, decreasing the average
KL-divergence by 94% compared to the Random baseline. The Joint approach solves
a joint optimization problem learning the most efficient followee configuration. Addi-
tionally, we show the impact of our approaches on the performances of the inference
pipeline using the precision at rank N (P@N) metric and, since we aim at suggesting as
few new followees as possible, the average amount of suggested profiles.
Finally, we test our concealing approaches in a more real world setting by evaluating
them against the Twitter’s Who To Follow [9] recommendation system. This system
recommends a target user other Twitter profiles to follow based (also) on his/her inferred
interests, which can be deduced from produced recommendations leveraging the same
DBpedia/Wikipedia alignments and categorical information used in the aforementioned
inference pipeline. We show how our approaches can partially equalize those inferred
interests, proving that our techniques have general applicability and can be used with
little or no knowledge about the target inference pipeline algorithm.
The rest of the paper is structured as follows. In Section 2, we briefly present the
related work. Then, we formally define our problem in Section 3, followed by the de-
scription of the user’s interest inference pipeline used as reference in Section 4. Our
concealing approaches are described in Section 5. Finally, we report the evaluation re-
sults in Section 6, closing with the conclusion in Section 7.
2 Related Work
User profiling Profiling of users in social media has been performed since their in-
ception both by the social media themselves and researchers. The inference of many
user profile attributes, such as gender [26], age [22], location [23], political affilia-
tion [26], level of education [11], and occupational class [21], has been studied both
for active [1,15,16,19,24,25] and passive users [3,7,20,26].
Followee information (i.e., the social graph) of a user plays a key role in user pro-
filing. One of the early studies successfully utilizing followee information to infer a
user profile attribute was Sadilek et al. [23]. The usage of GPS data of followees al-
lowed them to pinpoint the exact location (down to coordinates) of a target user with
80% accuracy. No additional information from the user’s profile was used. However,
their approach required a significant amount of high-quality location data to perform
inference, which is in many cases hard to acquire.
Zheleva and Getoor [26] exploited the social graph to infer gender and political
affiliation of users. They were able to profile both active and passive users. Additionally,
the more profiles of friends were public, the better the accuracy of such inference. Their
study is one of the first that raised important questions about the user privacy in the
social media. They concluded that the measures typically employed by social media to
protect personal data are not sufficient.
More recently, there has been an increasing interest in profiling interests of passive
users. Besel et al. [3] proposed an inference pipeline that utilize followee information
to infer interests. They link a target user’s followees to the corresponding Wikipedia
pages using their names and then category information is used to determine interests.
Our inference pipeline is constructed following their approach. Piao and Breslin [20]
iterated on this idea and produced a better performing system by improving the entity
linking step and using a different interest propagation approach.
In summary, social graph-based features have proven to be a useful feature in all
cases, confirming the idea that “you are what you follow”.
In most approaches for user profiling, a classifier is built for the inference of each
individual attribute. However, more holistic approaches were studied as well. For exam-
ple, Li et al. [12] proposed a two-layered structure. On the first layer they reimplemented
typical classifiers for attributes like location or education from previous works. Then
they built a Probabilistic Logical Reasoning framework and used the results from the
first layer as evidence. Since the accuracy of the first layer is not 100%, the second layer
should be able to account for possible errors and handle contradictory knowledge, effec-
tively preventing error propagation. Two reasoning approaches were explored: Markov
Logic Networks and Probabilistic Soft Logic. As a result they were able to work with a
wide variety of attributes: from gender to general like relationships towards entities.
Privacy in the social media Many user profiling studies cover the topic of users’ pri-
vacy in social media by warning of the potential risks of sharing private information in
user profiles. Privacy problems in social media, however, go beyond user profiling. Bet-
tini and Riboni [4], for instance, produced a comprehensive study on privacy preserva-
tion and the technological challenges arising in pervasive systems such as social media.
Felt and Evans [8] studied how social media themselves can protect users by re-
designing their APIs. They devised a privacy-by-proxy technique, where data is re-
vealed to applications only when needed, limiting the scope to prevent data harvesting.
Luo et al. [14] proposed to protect social media users by encrypting the user-related
information before it reaches the social media. Their approach aims at achieving a goal
similar to ours: not only to prevent the third parties from accessing the sensitive data of
the target user but also the social media themselves.
Kevin Ludlow [13] introduced the concept of Bayesian Flooding demonstrating
that the social media’s advertisement and recommendation systems can be confused
by flooding the user’s timeline with artificial actions.
Binarized Neural Networks Even though we did not employ neural networks in
this paper, our Joint approach was influenced by recent studies on Binarized Neural
Networks (BNN) [2,10]. We used the binarization and back-propagation procedures
from Hubara et al. [10] to find an optimal solution to our optimization problem.
3 Problem Definition
The goal of this paper is to protect the privacy of passive users by modifying their
lists of followees in such a way that makes it much harder for the target inference
pipeline to profile their interests. Followees are now being universally used by social
media as part of the digital footprint of a person and play a crucial role in inferring user
profile information such as interests. Even if the user does not post or share content on
the social media (passive user), followee data is still available to third parties and the
social network itself. The idea is to conceal this information without degrading a user
experience, which, in case of modifying his/her followee list, can be achieved by storing
the original unmodified list on a user-controlled device and use it to filter the timeline.
In this work we focus on concealing approaches tackling the inference of passive
users’ interests. Given a user u and his/her list of followees lu , a user’s interests inference
pipeline g(lu ) is designed to infer this user’s interests cu = g(lu ), cu ∈ Rn , cui ≥ 0, ∑i cui =
1, where n is the amount of interest categories and cui is the score of interest category i
for user u. The categories are then ranked based on their score cui and the final list of top
k categories is produced to represent the user’s interests. Real-world implementations
of such inference pipeline will have a set of thresholds to abstain from classifying a
user’s interests when their ranking is too ambiguous, i.e., cui scores of top categories are
very close to each other, making it impossible to reliably determine a user’s interests.
An ideal approach for concealing user interests would try to make all the interest
categories indistinguishable from each other. Therefore, the goal of our system is to
modify the information about the user, i.e., transform user profile u to user profile u0
(in our case produce a modified list of followees lu0 ), so that the target inference system
produces ambiguous results. Formally, the objective is to minimize the Kullback-Leibler
0
divergence3 between the cu = g(lu0 ) and the uniform distribution over possible interest
categories:
0 1 0 1 1 1 0
D(u0 ) = DKL (U{1, n}||cu ) = − ∑ log cui + ∑ log = − ∑ log cui − log n (1)
i n i n n i n
A possible limitation of such problem formulation is that the social network can
impose limitations on the amount of follow requests from the user and a large list of
followees can significantly increase the amount of API requests needed to acquire the
timeline thus creating a worse user experience. In this case we may require our system
to propose as few modifications to the initial followees list as possible and the final
objective will be as follows:
J(u0 , u) = (1 − α)D(u0 ) + α(|lu0 | −|lu |), α ∈ [0, 1) (2)
where α is a parameter that balances the tradeoff between minimizing the KL-divergence
(which requires adding followees to equalize inferred interest categories) and minimiz-
ing the amount of proposed followees.
4 Interests Inference Pipeline
To evaluate our concealing approaches (presented in Section 5) we have implemented
a user’s interests inference pipeline (g(lu ) in Section 3) that infers user’s interests based
on the list of followees. We follow Besel et al. [3] state-of-the-art approach, improving
it by removing dependencies from the Wikipedia API and the WiBi Taxonomy. We do
that by replacing the Entity Linking heuristics used there with our state-of-the-art re-
source called SocialLink [17,18] and pre-computing the category distributions over a
taxonomy of 49 top categories for all possible entities in English DBpedia/Wikipedia.
This enabled us to acquire a simple and robust system that allows testing different ap-
proaches for concealing a user’s digital footprint. More in detail, the pipeline employs
the following three-step procedure:
Fetch followees Followees lu of the target user u are fetched using the Twitter API.
3 http://en.wikipedia.org/wiki/Kullback-Leibler divergence
Link followees to DBpedia/Wikipedia Each followee profile f ∈ lu is linked us-
ing our state-of-the-art Linked Open Data resource called SocialLink. SocialLink con-
tains direct alignments between DBpedia entities for persons and organisations and
corresponding Twitter profiles. Each alignment in this resource is associated with a
confidence score s f that we use here to appropriately weight the contribution of each
followee f to the final user’s interest distribution. We want to make sure that our link-
ing procedure is robust since an error at this step will propagate along the pipeline.
Therefore, we selected a subset of m = 101, 769 high-quality alignments by setting cus-
tom conservative thresholds on confidence scores (with respect to the default ones used
in [17]) that provide 91% precision and 31% recall performance against SocialLink
gold standard.
Produce interest scores At this step each aligned followee has to be mapped to a
category distribution f, whose elements fi are the scores for interest categories i. Sim-
ilarly to [3], we use the DBpedia/Wikipedia category graph to propagate the specific
categories associated to the followee entity up to the 49 top-level categories here con-
sidered, for each of which a relevance score is computed. This process resulted in a
list of 3,507,016 scored DBpedia/Wikipedia entries. The interest scores in the category
distribution f are then normalized using a modified softmax function σ (f) to produce a
valid probability distribution across possible interests, where the normalized score σ (f)i
for interest category i is computed as follows:
(
z( fi ) ex if x 6= 0 ,
σ (f)i = n , z(x) = (3)
∑k=1 z( fk ) 0 if x = 0
This normalization procedure preserves zero scores for categories that were not ob-
served for the given followee f , thus reducing the noise across final user’s interests.
The interests distribution cu for the user u is finally computed as a weighted average of
normalized scores for user’s followees f ∈ lu :
∑ f ∈lu s f σ (f)
cu = (4)
∑ f ∈lu s f
The code of our interests inference pipeline along with the precomputed list of
scores for each aligned followee can be found in our GitHub repository.4
5 Concealing Approaches
In order to conceal a user’s interests we propose three approaches that calculate an
updated list of followees to minimize the objective defined by (2): Greedy approach,
Joint approach, and the baseline Random approach. In all cases, the system expects
the initial list of followees in input and chooses new followees from the same list of
pre-aligned profiles (from SocialLink) that our interests inference pipeline uses.
4 http://github.com/Remper/re-coding-ws
Fig. 2. Average KL-divergence for different Fig. 3. Average KL-divergence for different α
amounts of followees using Random approach values using Greedy approach
Random approach The most trivial direction we can take is to randomly follow new
people in hope that the category distribution will become closer to uniform. In this
approach, the profiles to follow are randomly drawn from the above-mentioned list and
the system will stop when 250 new unique followees are selected. Since the list has
its own bias towards certain categories, it can be expected that the more followees we
add this way the closer we get to the list’s own category distribution. This is why the
threshold on the amount of new followees has to be selected carefully in order to provide
a positive improvement in terms of our objective. Figure 2 shows how the average KL-
divergence changes based on the amount of followees proposed.
Greedy approach The greedy approach iteratively selects a new followee from the
pre-aligned list by picking the one that will decrease the KL-divergence between the
resulting category distribution and the uniform distribution the most. Therefore, it can
be seen as a breadth-first search over the space of possible configurations. The algorithm
stops when it is no longer possible to select a new profile to follow that will improve the
objective score (2). In our experiments the α parameter is set to 0.01. Figure 3 shows
how the average KL-divergence changes based on the choice of α.
Joint approach Finally, we have devised an approach that directly solves the formu-
lated optimization problem by jointly finding an optimal followee configuration. This
approach is inspired by recent studies about Binarized Neural Networks [10] and it ef-
fectively “learns” the binary mask wb where each element wbj corresponds to a possible
followee j to add, being 1 if that followee j should be followed and 0 otherwise.
Given the matrix of followees F ∈ Rm×n that is obtained by simply stacking row-
wise the category distributions σ (f) of the pre-aligned followee list (i.e., m = 101, 769)
following the normalization procedure from (3), the cu can be rewritten in terms of our
binary mask wb :
1
cu = wb F (5)
∑i wbi
The objective score can be computed as in (2). In order to solve this optimization
problem, we define an additional weight vector w ∈ Rm , and wb will now be computed
Fig. 4. An example histogram of user’s categories converging towards uniform distribution over
17k iterations of Joint approach. Each slice represent score distribution among categories for the
corresponding iteration. In a perfect scenario all scores should be equal to 1/n = 0.02.
using a simple deterministic binarization:
(
1 if w j ≥ 0 ,
wbj = Bin(w j ) = (6)
0 if w j < 0
Then w is learned by gradient descent towards the objective. Note that since the
derivative of the Bin function is zero almost everywhere, the gradient have to be back-
propagated using the straight-through estimator technique suggested by Bengio et al. [2].
The w parameters are initialized by drawing from N (2l, l), where l is the learning rate,
ensuring that the majority of possible followees are followed at the first iteration. Fig-
ure 4 shows how a user’s category distribution changes during the learning process.
6 Evaluation
We evaluate our concealing approaches Random, Greedy, and Joint against the user’s
interests inference pipeline of Section 4 and the Twitter’s Who To Follow recommen-
dation system. We measure four main performance metrics: (i) the average amount
of followee suggestions; (ii) the ability to conceal a user’s top-N interests categories,
measured taking the categories produced without concealing as gold standard and mea-
suring the precision P@N of the inference pipeline in reproducing those categories after
concealing is applied; (iii) the average delta between the first and the second category
probabilities; and (iv) the average KL-divergence between the category distribution pro-
duced by an approach and the uniform distribution across interest categories.
6.1 Evaluation Against Interest Inference Pipeline
In order to produce the dataset required to evaluate our concealing approaches, we gath-
ered a list of all the authors from ISWC proceedings for the last three years, extracting
Table 1. System evaluation against our interests inference pipeline
System performance (P@N) Score diff to Diff from uniform
Approach Suggestions
Top 5 Top 10 Top 15 2nd best (KL-div)
No mod 0 1.0 1.0 1.0 0.25 7.52
Random 250 0.39 0.58 0.65 0.12 3.15
Greedy 27.49 0.52 0.55 0.56 0.03 0.20
Joint 76.61 0.37 0.45 0.49 0.02 0.16
the names, titles and abstracts of all the papers associated with each author. After gath-
ering the list, the Social Media Toolkit [18] was used to find the corresponding Twit-
ter profile for each author, leveraging the collected textual context. This resulted in a
dataset of 491 Twitter accounts that were used to evaluate our approaches against the
user’s interests inference pipeline. The choice of the ISWC target audience was made
to showcase how easy it is to profile people in a particular community just by having a
set of names, keywords, or other related textual context.
The followees list for each author was gathered via the Twitter REST API and pro-
vided to our interest inference pipeline to produce an initial interest category distribu-
tion. Then, each approach was able to propose a modified list of followees and the cat-
egory distribution was recalculated. Table 1 shows the resulting performance for each
approach and the baseline numbers without applying any concealing technique. The
baseline is assumed to be perfect at predicting user interests (have 100% precision at 5,
10 and 15 top categories) since our goal is to hide true user interests from the inference
pipeline.
It can clearly be seen that, even though the Random approach already significantly
reduces the KL-divergence, it takes a significant amount of suggestions to achieve this
result. Random hides the initial top K categories of the user well, but produces a new
top category (usually, Sport) that stands out and makes the target inference pipeline
consistently producing false positives towards this category, which was not our goal.
The Greedy approach, on the other hand, produces an almost uniform distribution,
while providing a relatively small amount of suggestions. It does not hide the top K
categories as well as the Random approach, but the target system would have mostly
likely abstained to infer user’s interests given such category distribution. The results of
this approach clearly show that if the concealing system is able to predict the expected
score with high accuracy for an arbitrary followees list (in our case the approach had
perfect information), it is possible to confuse the target inference pipeline.
Finally, the Joint approach is able to find a more efficient followee configuration
than Greedy, producing the best results in almost all metrics at the cost of an increased
suggestions count. The amount of suggestions can further be tuned by setting different
values for the α parameter. However, we consider the current configuration to be a
reasonable tradeoff.
Table 2. System evaluation against Twitter’s Who To Follow
System performance (P@N) Score diff to Diff from uniform
Approach
Top 5 Top 10 Top 15 2nd best (KL-div)
No mod 1.0 1.0 1.0 0.84 12.23
Joint 0.20 0.37 0.47 0.44 8.56
6.2 Evaluation Against Twitter’s Who To Follow
In this scenario we evaluate the Joint approach, resulting the best in the previous eval-
uation, against the Who To Follow system used in production by Twitter to suggest
users to follow based on a target user’s profile [9].5 This system was chosen because
it provides a simple way to collect user’s interests as measured by Twitter. In order to
evaluate against this system, we have created a number of new Twitter users providing
as little initial information about the user as possible to the social network,6 apart the
lists of followees, so to simulate the behaviour of passive users.
After the creation of each account, the initial Twitter recommendations of users to
follow were gathered. These users represent popular Twitter accounts that are always
recommended to new Twitter users in a certain location. Then, a number of users were
followed from the pre-aligned list with the intention to give a clear bias towards some
interest category. After that, we gathered again the users to follow recommended by
the network. Finally, the Joint approach was used to propose the modified list and the
network’s suggestions were gathered one more time. Overall, three lists of Twitter user
recommendations were gathered, with the initial list acting as a filter to clean the other
two lists from the location-based and general popularity-based suggestions.
The remaining two lists were mapped to distributions of interest categories using
the formula presented in (4). Then the same metrics used in Section 6.1 were computed
to evaluate results. Unfortunately, since the Who To Follow box had to be gathered
manually and fresh Twitter accounts had to be created every time, the evaluation was
significantly downscaled compared to what we initially have hoped to measure. How-
ever, it can clearly be seen in Table 2 that, even though the concealing approach does
not have any information on how the target user’s interests inference system works, it
is often able to conceal the user’s true category distribution.
In order to achieve better results, a training set can be gathered to modify the fol-
lowee matrix F using the same manual gathering approach employed during the evalu-
ation. However, this is currently beyond the scope of this paper.
7 Conclusions and Future Work
In this paper we have presented approaches for concealing a social media user’s digital
footprint by proposing further people to follow, leveraging the alignment from Twitter
5 http://support.twitter.com/articles/227220
6 All user accounts were created using fresh email accounts using an IP address that can be
tracked down to Microsoft Azure cloud datacenter in Cheyenne, USA.
users and DBpedia/Wikipedia entries provided by SocialLink [17]. We showed that by
using simple techniques and without degrading user experience, passive Twitter users
can prevent the network or a third party system to infer their interests based on the
knowledge of who they follow. As our approach relies only on social graph information,
which is present in any social media, we believe it can be generalized and ported to other
platforms like Facebook and Instagram.
Even though the discussion about the privacy of users online has been a hot topic
lately, social media are reluctant to implement industry standard techniques such as
differential privacy and on-device computation, wanting instead to preserve their ability
to sell ads and promote their services. In this situation, we believe it is increasingly
important to explore various ways users can protect their digital identity.
In future work, further user profiling scenarios can be explored targeting different
attributes (e.g. location) and profiling use cases. We also believe that much can be done
to explore the privacy of active users, for example, making sure that the metadata shared
with the social media is limited. Regarding our use case, more sophisticated approaches
can be explored that are able to learn from the output of the interests inference pipeline
to produce more accurate results.
References
1. Abel, F., Gao, Q., Houben, G.J., Tao, K.: Analyzing user modeling on Twitter for personal-
ized news recommendations. In: Proc. of 19th Int. Conf. on User Modeling, Adaption, and
Personalization (UMAP). pp. 1–12. Springer-Verlag, Berlin, Heidelberg (2011)
2. Bengio, Y.: Estimating or propagating gradients through stochastic neurons. CoRR
abs/1305.2982 (2013)
3. Besel, C., Schlötterer, J., Granitzer, M.: Inferring semantic interest profiles from Twitter
followees: Does Twitter know better than your friends? In: Proc. of 31st Annual ACM Sym-
posium on Applied Computing (SAC). pp. 1152–1157. ACM (2016)
4. Bettini, C., Riboni, D.: Privacy protection in pervasive systems: State of the art and technical
challenges. Pervasive and Mobile Computing 17(Part B), 159 – 174 (2015), 10 years of
Pervasive Computing’ In Honor of Chatschik Bisdikian
5. Brunton, F., Nissenbaum, H.: Obfuscation: A user’s guide for privacy and protest. The MIT
Press (2015)
6. Dwork, C.: Differential privacy: A survey of results. In: Proc. of 5th Int. Conf. on Theory
and Applications of Models of Computation (TAMC). pp. 1–19. Springer-Verlag, Berlin,
Heidelberg (2008)
7. Faralli, S., Stilo, G., Velardi, P.: Recommendation of microblog users based on hierarchical
interest profiles. Social Network Analysis and Mining 5(1), 25 (2015)
8. Felt, A., Evans, D.: Privacy protection for social networking APIs (2008), http://www.cs.
virginia.edu/felt/privacybyproxy.pdf
9. Gupta, P., Goel, A., Lin, J., Sharma, A., Wang, D., Zadeh, R.: WTF: The Who to Follow
Service at Twitter. In: Proc. of 22nd Int. Conf. on World Wide Web (WWW). pp. 505–514.
ACM, New York, NY, USA (2013)
10. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized Neural Net-
works. In: Proc. of Advances in Neural Information Processing Systems (NIPS), pp. 4107–
4115. Curran Associates, Inc. (2016)
11. Li, J., Ritter, A., Hovy, E.H.: Weakly supervised user profile extraction from Twitter. In:
Proc. of 52nd Annual Meeting of the Association for Computational Linguistics (ACL). pp.
165–174. The Association for Computer Linguistics (2014)
12. Li, J., Ritter, A., Jurafsky, D.: Inferring user preferences by probabilistic logical reasoning
over social networks. CoRR abs/1411.2679 (2014)
13. Ludlow, K.: Bayesian flooding and Facebook manipulation (2012), http://www.kevinludlow.
com/blog/1610/Bayesian Flooding and Facebook Manipulation RD/
14. Luo, W., Xie, Q., Hengartner, U.: FaceCloak: An architecture for user privacy on social
networking sites. In: Proc. of Int. Conf. on Computational Science and Engineering. vol. 3,
pp. 26–33 (Aug 2009)
15. Michelson, M., Macskassy, S.A.: Discovering users’ topics of interest on Twitter: A first
look. In: Proc. of 4th Workshop on Analytics for Noisy Unstructured Text Data (AND). pp.
73–80. ACM, New York, NY, USA (2010)
16. Mislove, A., Viswanath, B., Gummadi, K.P., Druschel, P.: You are who you know: Inferring
user profiles in online social networks. In: Proc. of 3rd ACM Int. Conf. on Web Search and
Data Mining (WSDM). pp. 251–260. ACM, New York, NY, USA (2010)
17. Nechaev, Y., Corcoglioniti, F., Giuliano, C.: Linking knowledge bases to social media pro-
files. In: Proc. of 32nd ACM Symposium on Applied Computing (SAC). pp. 145–150. ACM
(2017)
18. Nechaev, Y., Corcoglioniti, F., Giuliano, C.: SocialLink: Linking DBpedia entities to cor-
responding Twitter accounts. In: Proc. of 16th Int. Semantic Web Conf. (ISWC). Springer
(2017)
19. Piao, G., Breslin, J.G.: Exploring dynamics and semantics of user interests for user model-
ing on Twitter for link recommendations. In: Proc. of 12th Int. Conf. on Semantic Systems
(SEMANTICS). pp. 81–88. ACM, New York, NY, USA (2016)
20. Piao, G., Breslin, J.G.: Inferring user interests for passive users on Twitter by leveraging
followee biographies. In: Proc. of 39th European Conference on IR Research (ECIR). pp.
122–133. Springer, Cham (2017)
21. Preotiuc-Pietro, D., Lampos, V., Aletras, N.: An analysis of the user occupational class
through Twitter content. In: Proc. of 53rd Annual Meeting of the Association for Computa-
tional Linguistics (ACL). pp. 1754–1764. The Association for Computer Linguistics (2015)
22. Rao, D., Yarowsky, D., Shreevats, A., Gupta, M.: Classifying latent user attributes in Twitter.
In: Proc. of 2nd Int. Workshop on Search and Mining User-generated Contents (SMUC). pp.
37–44. ACM, New York, NY, USA (2010)
23. Sadilek, A., Kautz, H., Bigham, J.P.: Finding your friends and following them to where you
are. In: Proc. of 5th ACM Int. Conf. on Web Search and Data Mining (WSDM). pp. 723–732.
ACM, New York, NY, USA (2012)
24. Siehndel, P., Kawase, R.: Twikime!: User profiles that make sense. In: Proc. of 11th Int.
Semantic Web Conf. (ISWC) – Posters & Demonstrations Track. vol. 914, pp. 61–64. CEUR-
WS.org, Aachen, Germany (2012)
25. Zarrinkalam, F., Fani, H., Bagheri, E., Kahani, M.: Inferring implicit topical interests on
Twitter. In: Proc. of 38th European Conf. on IR Research (ECIR). pp. 479–491. Springer,
Cham (2016)
26. Zheleva, E., Getoor, L.: To join or not to join: The illusion of privacy in social networks
with mixed public and private user profiles. In: Proc. of 18th Int. Conf. on World Wide Web
(WWW). pp. 531–540. ACM, New York, NY, USA (2009)