China Communications • December 2015
22
action between data objects. In particular, to
alleviate the cold-start problem, we learn the
effect of each location on individual users by
considering both his/her historical check-in
records and the geographic spatial distribution
of all POIs.
(2) We propose a unified and flexible rec-
ommendation framework based on a factor
graph model[12] to integrate different factors
(such as user preference and geo-social in-
fluence), which can automatically determine
the degree of each factor’s contribution by
optimizing the model parameters. Besides, we
propose a distributed learning algorithm to
improve the scalability of our method when
dealing with large-scale data sets.
(3) According to the two large-scale real
data sets collected from Foursquare and Gow-
alla, we conduct extensive experiments to
evaluate the recommendation accuracy and
scalability of our method. The experimental
results show that our method achieves better
performance than several state-of-the-art POI
recommendation approaches.
The remainder of this paper is organized as
follows. Section II presents an overview of the
related work, and Section III formalizes the
POI recommendation problem. In section IV,
we model the personalized two-dimensional
geographical influence using the data field
method; in Section V, we introduce a unified
geo-social recommendation framework in
detail; in Section VI, we discuss the experi-
mental results on the two real-world data sets.
Finally, Section VII summarizes this paper.
RELATED ORK
Recently, POI recommendation has emerged
as a popular topic in the eld of recommender
systems. According to the data available on
the Internet such as user check-in records,
GPS trajectory data[13], and text data[14], the
simplest way to provide a POI recommenda-
tion service is with those conventional recom-
mendation techniques such as collaborative
ltering. However, using only the information
of user check-in records is not enough to as-
terms of the geographic distance between lo-
cations. Prior studies assume that the distances
between POIs that a user visited before follow
a universal or personalized distribution[1, 2,
4-6], and the probability of a user visiting a
POI can be calculated according to the geo-
graphic distance between them. Unfortunately,
these studies are subject to some limitations,
e.g., the deciency of location’s intrinsic char-
acteristics and the difficulty in finding a rea-
sonable reference location[7]. Therefore, some
recent studies began to model the geographical
influence over two-dimensional geographic
coordinates (latitude and longitude)[7], aiming
at characterizing user check-in behavior better.
On the other hand, according to the social
correlation[8] of social theories, human move-
ment and mobility patterns are also affected by
their social friendships, and a few recommend-
er systems have utilized the social inuence in
conventional social networks to improve pre-
diction performance[9]. For example, accord-
ing to the result of our experiment conducted
on the data sets which were collected from
Foursquare[10] and Gowalla[5], although less
than 10% of a user’s check-ins are also visited
by his/her friends, the probability of checking
in the same POI for two friends is, on average,
much higher than that for two strangers.
To sum up, a unified framework integrat-
ing different factors seems to be a feasible
solution to improve the quality of POI recom-
mendation. How to model the personalized
geographical inuence in terms of geographic
spatial distribution, and further characterize
user check-in behavior better by integrating
the social inuence with the geographical in-
fluence, is the challenge we face. To address
the issue, in this paper we propose an effective
and scalable POI recommendation framework
that returns the top K POIs with the highest
scores to each target user. The main contribu-
tions of this paper are summarized as follows:
(1) We model the personalized geograph-
ical influence on user check-in behavior in
terms of geographic spatial distribution using
the data field method[11], because it has the
advantage in describing the non-contact inter-
This paper presents
a semi-supervised
probabilistic model
based on a factor
graph model to inte-
grate dierent factors.
To recommend more
appropriate POIs to
target users, both the
personalized geo-
graphical influence
and the social influ-
ence on individual us-
er’s check-in behavior
are considered.