International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

LOCATION PRIVACY PROTECTION MECHANISMS USING ORDER-

RETRIEVABLE ENCRYPTION FOR SOCIAL NETWORK
*1Ms. Shalini P, *2 Ms. Abarna N.

*1M.Phil Research Scholar, PG & Research Department of Computer Science & Information Technology, Arcot Sri
Mahalakshmi Womens College, Villapakkam, Tamil Nadu, India.
*2.Assistant Professor, PG & Research Department of Computer Science & Information Technology, Arcot Sri

Mahalakshmi Womens College, Villapakkam, Tamil Nadu, India.

---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract: A common functionality of many location- It does not only filter the noise, but also help to select
based social networking applications is a location sharing attractive and useful information. Recommender system
service that allows a group of friends to share their has achieved initial success based on a survey that shows
locations. With a potentially untrusted server, such a at least 20 percent of sales on Amazons website come
location sharing service may threaten the privacy of users. from the recommender system.Social networks gather
Existing solutions for Privacy-Preserving Location Sharing volumes of information contributed by users around the
Services (PPLSS) require a trusted third party that has world. This information is versatile. It always contains
access to the exact location of all users in the system or item/services descriptions (including textual descriptions,
rely on expensive algorithms or protocols in terms of logos and pictures), users comments, moods and users
computational or communication overhead. Other social circles, prices, and locations. It is very popular for
solutions can only provide approximate query answers. To recommending users favorite services from crowd-source
overcome these limitations, we propose a new encryption contributed information.
notion, called Order-Retrievable Encryption (ORE), for
PPLSS for social networking applications. The However, with the rapid increase in number of
distinguishing characteristics of our PPLSS are that it (1) registered Internet users and more and more new
allows a group of friends to share their exact locations products available for purchase online, the issue of cold
without the need of any third party or leaking any location start for users and sparsity of datasets has become
information to any server or users outside the group, (2) increasingly intractable. Fortunately, with the popularity
achieves low computational and communication cost by and rapid development of social networks, more and more
allowing users to receive the exact location of their friends users enjoy sharing their experiences, such as reviews,
without requiring any direct communication between ratings, photos and moods. The interpersonal
users or multiple rounds of communication between a relationships have become transparent and opened up as
user and a server, (3) provides efficient query processing more and more users share this information on social
by designing an index structure for our ORE scheme, (4) media websites such as Facebook, Twitter, Yelp, Douban,
supports dynamic location updates, and (5) provides Epinions, etc. The circles of friends also bring
personalized privacy protection within a group of friends opportunities and challenges for a recommender system
by specifying a maximum distance where a user is willing to solve the issues of cold start and sparsity.
to be located by his/her friends. Experimental results
show that the computational and communication cost of
our PPLSS is much better than the state-of-the-art
solution.

Keywords: Location privacy, location sharing services,

order-retrievable encryption, location-based social
networking, spatio-temporal query processing

I. INTRODUCTION
Recently people have been receiving more and
more digitized information from Internet, and the volume
of information is larger than any other point in time,
reaching a point of information overload. To solve this
problem, the recommender system has been created in Fig : Location based Services
response to the need to disseminate so much information.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 175
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

For the scope of this thesis we focus on Location- It apply the CI heuristic in a novel way. Typically,
Based Service (LBS). In these services the user's location is the CI heuristic is used to analyze how a new socio-
processed in order to deliver a service, such as finding a technical practice impacts the CI in very specific situations.
nearby Point-of-Interest (POI) or getting directions. Due to For example, consider a library that provides a paper-
their usefulness, LBS are being used by most mobile device based repository of all available books that can be used by
users. Some services require the users to reveal their visitors to locate the books in which they are interested. If
location either sporadically or continuously. The the library would replace this paper-based system by a
difference between the two is that in the sporadic case two computer-based search system, then the situation for the
consecutive queries to the LBS of the same user happen visitors would fundamentally change. While consulting the
with sufficient time in between so that they do not paper-based repository leaves no digital trace, any action
correlate. We shall understand LBSs from a broader on the computer-based system can be recorded. Finding
prospect than only from a technology perspective. Instead, the prevailing context of the CI heuristic is in this scenario
we consider processing location data as a new socio- relatively straight forward, because visitors use the
technical practice, having impact on technology, society system only in the library. This is different with LBSs as
and legislation. We adopt an interdisciplinary approach in they are employed in numerous situations and contexts. In
order to understand the threats on privacy from our work we analyze the impact of the new socio-technical
engineering, ethical and legal perspectives. practice on CI independent of a particular scenario.
Instead, we employ the context of travelling for our
For the domain of obfuscation-based location privacy analysis.
protection we first propose a framework that allows the
computation of optimal LPPMs for users who sporadically Location Data
engage in LBSs. Our framework accounts for resource
constraints of mobile devices with which LBSs are It presents the necessary background on location
typically being accessed. This makes our framework data. We provide a detailed explanation on apps that are
suitable to compute optimal dummy and precision-based the most important tools for people to utilize the
obfuscation strategies. An analysis of several optimal capabilities of their mobile devices, such as their location
obfuscation strategies further shows the trilateral trade- data. This also includes a description of the prevailing
off of privacy, Quality Loss (QL) and bandwidth overhead business model of the app-eco system and its privacy
issues, as well as the users' attitude towards this business
A concept of the rating schedule to represent user model. We continue with the explanation of the legal
daily rating behavior. protection of location data and outline users' attitude
The factor of interpersonal rating behavior towards LBSs. Finally, we provide a summary of the
diffusion to deep understand users rating threats to user privacy when location data is misused.
behaviors.
The main factors are, personal interest, Mobile Device Eco-System
interpersonal interest similarity, interpersonal
rating behavior similarity, and interpersonal Mobile devices are nowadays ubiquitous
rating behavior diffusion, into matrix factorization companions. More than two billion people worldwide are
with fully exploring user rating behaviors to using smartphones and more than 1.2 billion people are
predict user-service ratings. using tablet computers. Smartphones are intuitive to use
and they are equipped with a wide variety of
II. BACK GROUND communication modules, such as WiFi, Bluetooth and
4G/LTE, allowing their users to be connected in ways
never possible before. Another reason for the success of
Privacy assessment of LBSs from engineering, an
smartphones is their platform, i.e. the Mobile Operating
ethical and a legal perspective. Scholars from every of the
System (MOS) and the app-store that are maintained by
involved disciplines obtain an added value for their own
the Platform Operator (PO). More traditional mobile
field and additionally obtain an insight in how the topic is
phones usually run a MOS developed by the phone
addressed in other disciplines. For example, the detail
manufacturer and offered only a very limited number of
provided by the legal perspective serves as a valuable
additional applications, which typically have also been
reference for legal scholars. However, scholars from other
implemented by the phone manufacturer. The software of
disciplines may find the provided information especially
modern mobile devices, however, is much more
valuable, because works in their discipline typically lack
sophisticated.
this kind of legal detail. Also, the combination of the
engineering, ethical and legal discipline to assess privacy
implications in LBS contributes towards a better
understanding of privacy issues from a broader
perspective.
2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 176
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

Legal Protection of Location Data chooses a secret key, encrypts her location with the secret
key, encrypts the secret key with the public key of Bob and
There are several directives in place for the sends both to Bob. The receiver Bob uses his private key to
protection of location data of which we discuss in detail decrypt the symmetric key, enabling him to decrypt Alice's
the DPD, the ePrivacy Directive (2002/58/EC, as amended location. The data being sent between Alice and Bob could
by 2009/136/EC) (ePrivacy Directive) directive and the be transferred with the help of a central service provider,
recently published GDPR. The Data Protection Directive but if the service provider is reluctant to store and
(95/46/EC) (DPD) [190] defines the legal rules for forward encrypted data, Alice and Bob could exchange
processing personal data. The term processing is very their information with the help of a Distributed Hash
broad and can be anything from recording, handling and Table (DHT),
deleting data. The term personal data refers to "any
information relating to an identified or identifiable natural POI Finder
person (data subject)", whereas the identification can be
direct or indirect. Direct means to identify an individual One of the most commonly used LBSs allows users
without third party data sourcesMany models based on to find Point-of-Interest (POI) around their location.
social networks have been proposed to improve Therefore, the user submits a query to the service
recommender system performance. The concept of provider along with her location and, optionally, some
inferred trust circle based on circles of friends. To additional information on what kind of POIs the user is
recommend favorite and useful items to users. Their interested in. Private Information Retrieval (PIR) is suited
approach, called the Circle on Model, not only reduces the to implement a POI finder in a privacy-preserving way. PIR
load of big data and computation complexity, but also is a mechanism that allows users to query a database
defines the interpersonal trust in the complex social without the database server learning what information the
networks. user requested. It propose two protocols based on PIR.
The first protocol works with a single PIR request at the
The similarity between users or items according cost of providing only approximate results. The second
to the number of common ratings. An item-based CF protocol has a higher computational and communication
combined with a condition-based probability similarity overhead but provides more accurate query results. Both
and Cosine Similarity. Collaborative filtering-based protocols are built on a data structure based on Hilbert
recommendation approaches can be viewed as the first curves and search trees that convert the map coordinates
generation of recommender system. of POIs into 1-dimensional coordinates preserving the
proximity of POIs. This allows to apply PIR on originally
III. DESIGN OF PRIVATE LOCATION-BASED two-dimensional data.

Most LBSs for mobile devices are privacy invasive, Traffic Monitoring
because the service provider learns the user's location
data. A variety of services have been proposed that make The idea of traffic monitoring is that cars on the
this privacy invasion impossible. Such services employ road are equipped with tracking devices and report to a
cryptographic primitives that allow them to provide the LBS additional data, such as their current speed. This
necessary privacy guarantees. One of the most commonly allows the LBS to compute statistics, such as current road
used cryptographic primitives is homomorphic encryption utilization, that can be used to navigate cars in a more
that allows computations to be carried out on ciphertext. efficient way. Clearly, such a system can provide
There exists several cryptosystems that provide this considerable advantages, including the detection of traffic
homomorphic property. Cryptosystem possess the jams and the subsequent redirection of other cars to a
additive homomorphic property. Given two plaintexts m1 faster route. However, if cars constantly reveal their
and m2 and their respective encryptions E(m1) and E(m2), locations, any observer would be able to learn private
the following equation holds: information about the users

E.(mi) E.(m2) = E(mi + m2) -----(1) Formally define the Sybil detection problem.
Specifically, we first introduce the social Network model.
Geo-Social Networks Then we introduce a few design goals.

It proposes a private location-sharing service and

employs hybrid encryption to protect the user's privacy.
Every user is assumed to have a public/private key pair
and users can exchange their public keys either out-of-
band or with the help of the service provider. If Alice
wants to inform Bob about her current location, she

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 177
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

Tag assignments. The basic tagging behaviour,

namely tag assignment, is defined as a 3-tuple e: U
x / x T 6 {0,1}. If a user collected item t; with tag
t, then e t =1, otherwise e t =0.

User Profiling Based On Multidimensional Singular

Value Decomposition

Traditionally some two-dimensional CF

approaches apply SVD on a user-item matrix to compute
user or item profiles and identify similar users/items
(Symeonidis et al., 2006; Mi Zhang & Hurley, 2009). In
these approaches, user-item matrix M 6 R't/'x'7' is
decomposed and approximated by the truncated SVD:

Taking user-based CF, for example, <W|[/|XFC 1

Fig : The propagation in Sybil fcxfc is used to project each user's data from an |/|-
dimensional space to a k-dimensional space, where k
Let us consider an undirected social network G = principal components of the data are preserved. Thus the
(V;E), where a node v 2 V represents a user in the system user profile matrix in the SVD-based user profiling is
and an edge (u; v) 2 E indicates that the users u 2 V and v 2 computed as:
V are socially connected. In an ideal setting, G represents a
weighted network of trust relationships between Users, Recall that we have a social network G = (V, E) of
where the edge weights represent the levels of trust the nodes in the system. Each node can have two states,
between users [92]. Each node is either benign or Sybil. i.e., benign or Sybil. Thus, we associate a binary random
Figure 3.1 illustrates a Sybil attack. We denote the sub variable * +1g with each node. xv = +1 means that
network including the benign nodes and the edges node v is a benign node and xv = 1 indicates that node vis
between them as the benign region, denote the sub Sybil. In the following, we use xA to represent the set of
network including the Sybils and edges between them as random variables associated with the nodes in the set A.
the Sybil region, and denote the edges between the two Moreover, we use _xA to denote the observed values of
regions as attack edges. Note that the benign region could these random variables. There might exist some prior
consist of multiple communities and we will evaluate their information about a node v independently from all other
impact on Sybil detections nodes in the system. Such prior information could be the
content generated by v or its behavior. We model the prior
IV. PROPOSED ANALYSIS belief of v being benign as follows:

In our research, user, item and tag are three ( ) ,

( )
fundamental entities of data. For the three basic entities,
we define {a/, b/ and T as disjoint non-empty finite sets. Where hv quantities the prior information about v. More
We use set symbols in lower case with indices in the specifically, hv> 0 encodes the scenario in which v is more
subscripts to denote individual elements of a set, for likely to be benign; hv< 0 encodes the opposite scenario;
example, u2 denotes the second user in the user set {c/. We hv = 0 means prior information is not helpful to determine
use symbols of individual set elements, i.e., user , item iy, v's state. here now introduce * ( ) + , the set of v's
or tag tp, in the subscripts or superscripts of the set neighbors in the social network, and their respective
symbols, as conditions imposed onto the sets to denote states _xv . When these states are known, the probability
subsets. For example, denotes the item set in which each of v to be benign is modeled as
item was used by the vth user u 6 U.

Users. U = u2, , u^, ,U|y|} contains all users in ( )

(
a social tagging system. denotes the vth user, u56
0, 1 <v < |a[a/b|. A Pair wise Markov Random Field
Items. / = (t1,t2, , ty, , t|7|} contains all items
used by the users in the system. iy denotes the We find that the probabilistic local rule
y'th item, iy 6 /, 1 < _/' < |a/b|. introduced in the previous section can be applied by
Tags. T = (t1, t2, , tp, , t|T|} contains all tags modeling the social network as a pair wise Markov
used by the users in the system. tp denotes the pth Random Field (MRF). A MRF defines a joint probability
tag, tp 6 T, 1 < p < |T|.
2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 178
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

distribution for binary random variables associated with

all the nodes in the network. Specifically, a MRF is
specified with a node potential for each node v, which
incorporates prior knowledge about v, and with an edge
potential for each edge (u; v), which represents
correlations between u and v. In the context of Sybil
detection, we define a node potential _v(xv) for the node v
as

( ) {

Fig : Order Retrieve Encryption Work Process

As mentioned above, user locations (i.e., points) in

the database server are always in encrypted form. When
an encrypted query location of Q1 for a group of friends
is received by the database server, the database server
should determine for any two friends encrypted locations
within the group which of them is closer to the encrypted
query location. To achieve this, we use our proposed
encryption notion ORE for geographical data. An ORE
scheme is a symmetric key encryption scheme with two
additional functions: one is for generating encrypted
query locations and the other one is for the database
server to determine which one between two encrypted
user locations is closer to an encrypted query location. The
scheme is called ORE because the order of the encrypted
user locations in terms of their distances from any given
encrypted query location can be retrieved. Note that the
Fig : Markov Random Model Implementation actual distance information is not retrievable.

Algorithm Implementation

1. Rating scale of individual users

2. Popularity of individual items
3. User preferences
rui(t)= +bu(t)+bi(t)+qTipu(t)

Parameterizing the model

rui(t)= +bu(t)+bi(t)+qTipu(t)

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 179
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

Use functional forms: bu(t)=f(u,t), bi(t)=g(i,t), Hotels & Travel, Night Life, Pets, Restaurants, and
pu(t)=h(u,t) Shopping.
Need to find adequate f(), g(), h()
General guidelines: LOCAL TRUSTEE SELECTION STRATEGIES
Items show slower temporal changes
Users exhibit frequent and sudden changes For a user u, a local trustee selection strategy
Factors pu(t) are expensive to model essentially computes a score s(v, u) for each friend v of u
Gain flexibility by heavily parametering the and then selects my friends with the highest scores as u's
functions trustees.

1. User Rating Behavior Exploration T-Random: As a baseline, T-Random assigns a random

number ranging from 0 to 1 as the score s(v, u).
The factors of interpersonal interest similarity
Wu,v and personal interest Qu,i proposed in have been T-CF: As was shown by the number of common friends of
proved effective. Thus, in this subsection, we turn to the two users is an informative indicator about the level of
details of our proposed interpersonal rating behavior trust between them. Thus, one speculation is that a user
similarity and interpersonal rating behavior diffusion. might select friends with which he or she shares many
common friends as trustees. To quantify the security of
2. Interpersonal Rating Behavior Similarity: this speculation, we design the strategy T-CF (i.e.,T-
Common Friends), which uses the number of common
The behavior habit is essential. It could not be friends shared by u and his or her friend v as the score
separated from temporal information. Thus, we define
rating behavior in this paper as what the user has done ( ) , ( ) ( )-
and when it happened. This kind of behavior presentation
However, there are two drawbacks of T-CF. First, the fact
arouses us to the curriculum schedule. The schedule
that u shares many friends with a popular user v doesn't
arranges which course would we take and when we
necessarily mean that u and v have a high level of trust
should go to class.
because it is normal for many friends of u to be in v's
friend list. Second, if a common friend of u and v is a
3. Interpersonal Rating Behavior Diffusion:
popular user, then sharing him or her doesn'tnecessarily
indicate a high level of trust between u and v. Next, we
We consider the factor of social users rating
introduce two strategies to overcome the two drawbacks,
behavior diffusions. We explore the diffusion of user rating
respectively.
behavior by combining the scope of users social network
and the temporal information of rating behaviors. For a
T-JC: To overcome the first drawback of T-CF, we design
user, we split his/her social network into three
the trustee selection strategy T-JC (i.e., T-Jacquard
components, direct friends, mutual friends, and the
Coefficient), which uses the Jacquard Coefficient [61] of
indirect friends.
the two sets T(u) and T(v) as the score ( ) i.e.,
4. Impact of Predicted Integer Ratings:
( ) ( )
The predicted ratings are decimal; we discuss the ( ) ( ) . (2)
impact of predicted integer ratings on performance. The ( ) ( )
ratings user rated are all discrete values ranging from 1 to
5. But the predicted results of matrix factorization model T-AA: To overcome the second drawback, we design the T-
are all decimals. Thus, it is necessary to discuss the impact AA (i.e., T-Adamic Ada) strategy, which uses Adamic-Ada
of integer predicted ratings. We round decimal ratings we similarity between u and v as the score s(v,u). Adamic-
predicted into discrete integers. Ada similarity penalizes each common friend by its
popularity (i.e., the number of friends). Formally, we have
5. Yelp Dataset:
( ) ( ) ( ) ( )
Yelp is a local directory service with social
networks and user reviews. It is the largest review site in T-Degree:, users could be those having large out degrees
America. Users rate the businesses, submit comments, in the trustee network, and they could enable an attacker
communicate experience, etc. It combines local reviews to compromise many other users. Thus, we propose the T-
and social networking functionality to create a local online Degree strategy to minimize the maximum out degree in
community. Yelp dataset4 contains eight categories, the trustee network. Intuitively, T-Degree constrains that
including Active Life, Beauty & Spas, Home Services, no users are selected as trustees by too many other users.

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 180
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

Algorithm 2 shows our T-Degree strategy. T-Degree

selects trustees for users one by one. For each user u that
has adopted the trustee-based social authentication
service, T-Degree selects his or her m u friends whos
current out degrees in the trustee network are the
smallest as u's trustees; ties are broken uniform at
random.

V. EVALUATION RESULT

The PLQP is a suit of protocols supporting Fig : The comparison of each entitys time cost
privacy-preserving LBS in mobile application. It has high between our scheme and PLQP
efficiency and achieves fine-grained control by exerting
the CP-ABE scheme, which is similar to our work. Thus, in
this section, we will compare our scheme with the PLQP
scheme for evaluating the performance of our proposed
scheme. The algorithms are implemented using the Big
Integer library on a Windows 8.1 system with Intel CORE
i7-4500U [email protected] GHz and 8.00 G RAM. We have 10
tests in this experiment. Additionally, in each test, we use
1000 pairs of random locations for the publisher and the
querier, respectively. We present the average results for
each test in the following figures.It shows the detailed
time cost for once location distance compute and location Fig: The time comparison of performing
distance compare, respectively. It is obvious that the time location distance compute between our scheme and
cost at the publisher is always zero, PLQP

Fig: The time comparison of performing

Fig : distance compute and location distance compare location distance compare between our scheme and
PLQP
The comparison results of the total time cost for the
aforementioned two operations, respectively. It can be In our system, the publisher can authorize the
seen that our scheme is much more efficient than PLQP. queries that he/she knows or not, such as his/her friends
shows the comparison results of the detailed time cost for or someone who has similar interests. Hence, the queries
the aforementioned two operations at the publisher, may include attackers. If so, it is easy for the attacker to
querier and mobile cloud, respectively. To be more clear, get a certain plaintext/ciphertext pair. Thus, our scheme
we also present the comparison results in. From these has to be secure against the chosen plaintext attack. Next,
figures, we get three points: we will prove it

At the querier, the time cost in our scheme is much Conclusion

less that that in PLQP.
At the publisher, the time cost in our scheme is zero. In this Paper , here propose a user-service rating
At the mobile cloud, the time cost in PLQP is zero prediction approach by exploring users rating behaviors
with considering four social network factors: user
personal interest (related to user and the items topics),
interpersonal interest similarity (related to user interest),

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 181
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 08 | Aug -2017 www.irjet.net p-ISSN: 2395-0072

interpersonal rating behavior similarity (related to users circle, IEEE Trans. Knowl. Data Eng., vol. 26, no. 7, pp.
rating habits), and interpersonal rating behavior diffusion 17631777, Jul. 2014.
(related to users behavior diffusions). A concept of the 10. K. Lee and K. Lee, Using dynamically promoted
rating schedule is proposed to represent user daily rating experts for music recommendation, IEEE Trans.
behavior. The similarity between user rating schedules is Multimedia, vol. 16, no. 5, pp. 12011210, Aug. 2014.
utilized to represent interpersonal rating behavior 11. Z. Wang et al., Joint social and content
similarity. The factor of interpersonal rating behavior recommendation for user generated videos in online
diffusion is proposed to deep understand users rating social network, IEEE Trans. Multimedia, vol. 15, no. 3,
behaviors. We explore the users social circle, and split the pp. 698709, Apr. 2013.
social network into three components, direct friends,
mutual friends, and the indirect friends, to deep
understand social users rating behavior diffusions. These
factors are fused together to improve the accuracy and
applicability of predictions. We conduct a series of
experiments in Yelp and Douban Movie datasets. The
experimental results of our model show significant
improvement.

Future work

To further enhance the recommendation, check-in

behaviors of users will be deeply explored by considering
the factor of their multi-activity centers and the attribute
of POIs.

REFERENCES

1. K. Mao, L. Shou, J. Fan, G. Chen, and M. Kankanhalli,

Competencebased song recommendation: Matching
songs to ones singing skill, IEEE Trans. Multimedia,
vol. 17, no. 3, pp. 396408, Mar. 2015.
2. R. Keshavan, A. Montanari, and S. Oh, Matrix
completion from noisy entries, J. Mach. Learn. Res.,
vol. 11, pp. 20572078, 2010.
3. Q. Liu, E. Chen, H. Xiong, C. Ding, and J. Chen,
Enhancing collaborative filtering by user interest
expansion via personalized ranking, IEEE Trans. Syst.,
Man, Cybern. B, Cybern., vol. 42, no. 1, pp. 218233,
Feb. 2012.
4. Y. Chen and J. Canny, Recommending ephemeral
items at web scale, in Proc. SIGIR, 2011, pp. 1013
1022.
5. M. Harvey, M. J. Carman, I. Ruthven, and F. Crestani,
Bayesian latent variable models for collaborative
item rating prediction, in Proc. CIKM, 2011, pp. 699
708.
6. X.-W.Yang, H. Steck, andY. Liu, Circle-based
recommendation in online social networks, in Proc.
KDD, 2012, pp. 12671275.
7. Y. Zhou and L. Liu, Social influence based clustering of
heterogeneous information networks, in Proc. KDD,
2013, pp. 338346.
8. H. Gao, J. Tang, X. Hu, and H. Liu, Exploring temporal
effects for location recommendation on location-based
social networks, in Proc. RecSys, 2013, pp. 93100.
9. X. Qian, H. Feng, G. Zhao, and T. Mei, Personalized
recommendation combining user interest and social

2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 182