Documente Academic
Documente Profesional
Documente Cultură
D.Hemamalini
Abstract - Recommendation systems have received considerable attention in the recent decades. Yet with the
development of information technology and social media, the risk in revealing private data to service providers
has been a growing concern to more and more users. With the rapid development of global position system and
location-based social networks (LBNS) have attracted millions of users to share rich information. Point-of-
interest (POI) recommender system plays an important role in LBSNs. Data collected through LBSN activity can
enable better recommendations of places or points of interest such restaurants and malls. Point-of-Interest
recommendation approaches from traditional recommendation approaches. The decision process for a user to
choose POI is complex and can be influenced by numerous factors, such as personal preferences, geographical
considerations, and user mobility behaviors. Geographical proximity, known as a unique feature of LBSNs,
significantly affects user check-in behavior. However, most of prior studies characterize the geographical
influence based on a universal or personalized distribution of geographic distance, leading to unsatisfactory
recommendation results. In this paper, design a framework for the personalized geographical influence in a two-
dimensional geographical space is modeled using the data field method, and we propose a semi-supervised
probabilistic model based on a factor graph model to integrate different factors such as the geographical
influence. Moreover, a distributed learning algorithm is used to scale up our method to large-scale data sets.
Experimental results show that our method outperforms other competing POI recommendation techniques.
Index Term Recommender system, Geographical Proximity, LBSN, POI, Distributed Learning Algorithm
Unlike the traditional recommender sys- tems that (1) We model the personalized geograph- ical
focus on non-spatial items (such as books and influence on user check-in behavior in terms of
movies), POI recommendation requires users to geographic spatial distribution using the data field
visit or check in locations in a LBSN. Geographical method[11], because it has the advantage in
proximity has a great impact on user check-in describing the non-contact inter-action between data
behavior[2], that is, nearby friends always share objects. In particular, to alleviate the cold-start
more common POIs than the others who live far problem, we learn the effect of each location on
away[3]. However, human mobility behavior individual users by considering both his/her
presents a challenge for the methods that make historical check-in records and the geographic
recom- mendations based on the distances between spatial distribution of all POIs.
users places of residence. To address the issue,
(2) We propose a unified and flexible rec- have studied social recommendations solely based on
ommendation framework based on a factor graph a social graph. Their results suggest accurate
model[12] to integrate different factors (such as user recommendations inevitably leak information about
preference and geo-social in- fluence), which can
automatically determine the degree of each factor the existence of edges between certain nodes. Our
s contribution by optimizing the model parameters. work differs from theirs because a) we are trying to
Besides, we propose a distributed learning hide the social relations from the SP and b) social
algorithm to improve the scalability of our method recommendations in our framework are integrated
when dealing with large-scale data sets. with other recommender systems. The closest
references to our research can be found in the
(3) According to the two large-scale real data sets literature of privacy for social relations in social
collected from Foursquare and Gow- alla, we networks. Anderson et al. [3] have proposed an
conduct extensive experiments to evaluate the architecture for social networking that protects users
recommendation accuracy and scalability of our social information from both the operator and other
method. The experimental results show that our network users. Domingo-Ferrer et al. [15] have
method achieves better performance than several proposed a scheme to prevent the resource owner in
state-of-the-art POI recommendation approaches. social networks from learning the relationships and
trust levels between the users who collaborate in the
II. RELATED WORK resource access. Tootoonchian et al. [32] have
proposed a cryptographic framework, Lockr, to
Recommender systems utilizing the information separate social networking content from all other
shared in social networks have recently attracted online social network functions. Backes et al. [4]
tremendous attention from both academia [21, 23, 24, have proposed a similar approach to achieve access
33] and industry [34]. Konstas et al. [21] have control while keeping social relations hidden from
proposed a recommendation algorithm that performs the SP. We exploit a similar idea as relation
a Random Walk with Restarts on a graph established authentication [4] to implement access control.
among users, items and tags. Ma et al. [23, 24] have
incorporated users social friendships into model- Recently, POI recommendation has emerged as a
based collaborative recommender systems. All these popular topic in the field of recommender systems.
approaches aim to recommend traditional items, such According to the data available on the Internet
as movies; the effectiveness of these approaches in such as user check-in records, GPS trajectory
geo social networks has not yet been evaluated. Ye et data[13], and text data[14], the simplest way to
al. [33] have proposed a unified POI recommendation provide a POI recommenda- tion service is with
framework in geosocial networks, which fuses user those conventional recom- mendation techniques
preferences for POI with social influence and such as collaborative filtering. However, using only
geographical influence. the information of user check-in records is not
enough to as- sure accurate POI recommendation
Some of the earliest privacy-preserving approaches results, which suggests that more additional
for recommender systems come from Canny [7, 8], informa- tion should be considered together.
who has proposed using homomorphic encryption
and a peer-to-peer protocol to provide privacy for The geographical feature of POIs is actual- ly a
several model-based collaborative recommender unique identifier distinguishing them from other non-
systems. Similar ideas have been explored [17,26]. spatial items[7]. On the one hand, some prior
Polat and Du [29] have proposed that customers use a studies[3] calculated the similarity between users in
randomized perturbation technique to disturb their terms of their geographical influence, and then
private data before sending the data to the SP. Similar recommended possible POIs using collaborative
ideas have also been explored [5, 20]. Ameur et al. filtering techniques. On the other hand, other prior
[2] have proposed using a semi trusted third party to studies investi- gated the geographical influence of
distill encoded sensitive customer information. locations. For example, Cheng[15] et al. viewed
loca- tions as ordinary non-spatial items and quan-
Previous work aims to keep individual user records tified the geographical influence of locations by
hidden from the SP or other users. Our research predefining a range; Kurashima[16] et al. proposed
differs from these approaches because our research a geo-topic model based on the as- sumption that if a
aims to keep users social relations secret from the location is closer to a users current location or the
SP, while still providing users with the benefits of locations he/she visit- ed, it is more likely to be
social recommendations. Machanavajjhala et al. [25] visited by the user; Lian[17] et al. proposed a
weighted matrix factorization model which check-in variable) representing the frequency of
augmented users and POIs latent factors in the check- ing in POI vj (with lj) by user ui. Assuming
model with ac- tivity area vectors of users and
influence area vectors of POIs, respectively. that CL is an observed check-in action matrix, the
set of observations can be denoted as O = (U; V; G;
In many prior studies[1, 2, 4, 5], the dis- tances CL). According to all the observations, our task is to
between locations visited by the same user were predict the probability that a giv- en user visits a new
modeled based on a common distri- bution for all POI, and to recommend the top-k POIs to the user.
users, e.g., a power-law distri- bution or a multi-
center Gaussian distribution. Although this benefits IV GEOGRAPHICAL INFLUENCE
the quality of POI recommendation to some extent,
a common distribution for all users often ignores the According to the Toblers first law of geog-
per- sonalized geographical influence. Therefore, raphy[20], the locations near the place where a user
Zhang[6] et al. utilized kernel density estima- tion to usually visits, are more likely to be recommended
estimate the distribution of distances between pairs to the user. Users differ in the geographic spatial
of locations for each individual user. Furthermore, a distribution of already-vis- ited locations. The
latest study[7] attempted to characterize user check- frequency distribution of a users historical check-
in behavior using two-dimensional kernel density ins reflects his/her preference for different locations,
estimation. and the preferences of various users for the same POI
are also different. Therefore, modeling the geo-
Inspired by the findings in traditional social graphical influence of locations on user check- in
networks, several POI recommendation ap- behavior is the key to solving the problem in this
proaches[3, 5, 14] took into consideration the social paper. Unlike the statistical method used in the
influence, but the results indicated that the effect of latest study[7], in this paper we use the data field
such a factor was not significant, because there was, method[11] to model the mutual interaction between
on average, a small over- lap of a user s check-ins locations by constructing virtual fields in a two-
to his/her friends check-ins. Besides this one, dimensional geographi- cal space.
some researchers have also utilized the information
of time and content to provide better This study assumes that the location of a POI
recommendation re- sults. For example, Gao[18] et represents a two-dimensional data point and that the
al. studied the temporal cyclic patterns of check-ins number of user check-ins for the location is the mass
in terms of temporal non-uniformness and of the data point. A data field in the two-
temporal consecutiveness, and they proposed a dimensional geographic spa- tial space is then
time- aware POI recommendation model; Yin[19] constructed. We illustrate the modeling process with
et al. proposed a location-content-aware topic model a simple example.
in consideration of both personal inter- est and local
preference. Figure 1 shows the distributions of two users check-
in locations, and Figure 2 further shows the
corresponding distributions of personal check-in
III. PROBLEM DEFINITION
locations in Figure 1 in terms of data field. In each
data field, we can calculate the potential value of a
We explore the information of user preference, social
given point, which is regarded as the geographical
influence and personalized geographical influence,
influence of a location on an individual user. The
and then integrate them into a uni- fied probabilistic
higher the potential value, the more likely a user is to
model to recommend proper POIs for target users.
prefer the corresponding location.
Next, we formalize the problem of this study.
5.3 Performance
i=1.M
POI cj in terms of user interest topic distribution and POI topic distribution j . The second term Pj is
a regional level popularity factor for POI cj as a word-of-mouth opinion on the POI. is a factor to balance these two
factors. Then TLij considers both interest topic match between user and POI, and location aware word-of-mouth
opinions for a POI.