2017 6th IIAI International Congress on Advanced Applied Informatics

A Point-of-Interest Recommendation Method Using

Location Similarity

Jun Zeng Yinghua Li

Key Laboratory of Dependable Service Computing in Graduate School of Software Engineering,
Cyber Physical Society, Chongqing University,
Chongqing University, Chongqing, China
Chongqing, China [email protected]
[email protected]

Feng Li Junhao Wen Sachio Hirokawa

Graduate School of Software Graduate School of Software Research Institute for Information
Engineering, Engineering, Technology,
Chongqing University, Chongqing University, Kyushu University,
Chongqing, China Chongqing, China Fukuoka, Japan
lifeng@ cqu.edu.cn [email protected] [email protected]

Abstract—POI recommendation aims to recommend places data. Collecting a certain scale of users’ check-in location
which users have not visited before. In this paper, we proposed and analyze these data can find users’ interests and activity
a POI recommendation method using location similarity, rules. For example, a user will go to a certain restaurant to
which assumes that people may be interested in the places that have lunch, swim in the afternoon and watch movie in the
are similar with the places that they have been to before. In evening, etc. These user’ behaviors, potential rules and
order to calculate the similarity of locations, we proposed a hobbies can help recommender system to recommend
novel method using time slots. Every two hours can be movies, restaurant, hotels, tourist attractions.
considered as a time slot. In other words, one day can be
segmented into 12 time slots. For each location, the check-in In this paper, we analyze check-ins from social network
times in each time slot can be collected. These check-in times and recommend Point-of-Interest (POI) for users. POI
can form a vector, which can be used to calculate the similarity recommendation aims to recommend places which users
of two locations. According to the similarity, the score of each have not visited before. Some researchers have done massive
unvisited locations can be calculated and sorted. Finally, the researches in the POI recommendation system. Collaborative
POI recommendation can be generated from the top-n filtering (CF) is widely applied to POI recommendation
unvisited locations. The experiment results show that the because of simplicity and effectiveness. There are much
proposed method is effective. work about memory-based CF, user-based CF and model-
based CF. However, available information has not been fully
Keywords—Time slot; POI recommendation; Location
utilized in previous work. Most of them only consider
similarity; Check-in data.
limited factors when recommending POIs to users.
I. INTRODUCTION In this paper, we proposed a POI recommendation
method using location similarity, which can also be
In recent years, smartphone has been an inalienable part considered as a location-based CF. This method assumes that
of people’s life. People can obtain all the information they people may be interested in the places that are similar with
need at their fingertips every day. The social media, such as the places that they have been to before. In order to calculate
Facebook, Foursquare, Twitter, etc., is an important function the similarity of locations, we proposed a novel method
of smartphone. With a social application people can engage using time slots. Every two hours can be considered as a
with friends and family, post new photos, update their status, time slot. In other words, one day (from 0:00-24:00) can be
and defeat boredom. The social software is convenient for segmented into 12 time slots. For each location, the check-in
users to search restaurants, hotels, attractions, places of times in each time slot can be collected. These check-in
entertainment in user’s surrounding area. Many social times can form a vector, which can be used to calculate the
platforms have more than tens of millions users. When users similarity of two locations.
comment or update photos and videos on these social
networks, users’ location information can be acquired from
users’ behavior. These data are so-called online check-in

978-0-7695-6178-3/17 $31.00 © 2017 IEEE

978-1-5386-0621-6/17 435
436
DOI 10.1109/IIAI-AAI.2017.122
Authorized licensed use limited to: Fernuniversitaet Hagen. Downloaded on December 13,2023 at 14:05:14 UTC from IEEE Xplore. Restrictions apply.
 The rest of this paper is organized as follows. In Section Yin, et al. [6] proposed the user interest drift will make
Ċ, we introduce related work. In Section ċ, we describe interest changed when users are in different places. Paper [6]
dataset and time slot. In Section Č , we describe new mainly considered the user's interest from different locations
recommendation method incorporating time slot influence of the theme, setting up two models in the local and foreign
preferences according to the people.
and location similarity. Section č presents experiment
result. Finally, section Ď concludes this paper.

II. RELATED WORK III. PRELIMINARIES

In recent years, recommendation technology [1-15] has A. DataSet
attracted the attention of scholars. POI Recommendation in
LBSN has received more and more researchersÿ attention In this paper, 194108 Foursquare check-in data, which
[1-11]. were made within Singapore between Aug.2010 and
Jul.2011, are chosen as the Dataset. The dataset contains
Collaborative Filtering’s characteristics is based on the 2321 users and 5596 locations. Each data include user ID,
history of the user attendance records, looking for users with location ID, geographical coordinate, check-in time and date
the similar preferences, then recommends places to users. ID. TABLE I shows the samples of the dataset.
Mayy habayeb, et al. [4] presented through the role of the
similar situs trajectories and network user to provide users
with recommendations, and the biggest difficulty of the TABLE I. THE SAMPLES OF THE DATASET
location trajectories is sparse historical trajectory data, and
the interaction of web users can find more interesting Check-in Date
User ID Location ID Coordinate
Time ID
information, add this information to the track. Context based
USER_1675 LOC_1967 (1.31, 103) 00:48 0
on Recommendation. It characterized in that using the
USER_1544 LOC_2505 (1.34, 103) 02:16 0
information of item, finding the features of item and
USER_855 LOC_3369 (1.29, 103) 18:55 14
analyzing the preferences of user to recommend. Huiji Gao, USER_855 LOC_2909 (1.28, 102) 17:57 13
et al. [3] presented a unified framework where the theme (the … … … … …
site is about what), the interest of users (whether the user is
interested in this space), the display of view (the site how),
the relationship between the user location and sign, and
unifying three directory information into a framework, In Foursquare dataset, one user may check-in at different
investigating each directory information in the real word. places at different time. Similarly, one place can be
This method can effectively solve the data sparseness checked-in by different users.
problem.
B. Time slot
Knowledge-based recommendation [16] is a derivation From TABLE I, it should be noted that the distribution of
method based on a certain rule, which can guide the user to check-in time is diffuse. In order to make the analysis easy,
find the items of interest in a large number of options. The we introduce time slot to split one day into 12 time slots,
advantage of this method is that it can have a which is shown in Fig. 1.
recommendation without rating data, and the disadvantage is
that the specific rules are difficult to institute.
Hybrid recommendation [17] method utilized the
advantages of the former three methods, and overcome the
disadvantages of the above mentioned algorithm. This
method is a combination of various recommendation
algorithms which is widely used in the recommendation
system.
In addition, paper [2] established a model based on
interest and geographical distance. Literature [5] proposed a
many-sided personalized recommendation system. Hongzhi Fig. 1. Splitting one day into 12 time slots

This research is supported by Supported by the National Natural After the time slots are determined, the check-in time can
Science Foundation of China (Grant No. 61502062 and Grant No. be simplified. For example, if the check-in time is 00:48, it
61672117), the China Postdoctoral Science Foundation under Grant
2014M560704, the Scientific Research Foundation for the Returned
can be denoted by t1. Similarly, if the check-in time is 01:54,
Overseas Chinese Scholars (State Education Ministry), the Fundamental it can be also denoted by t1. Although, 00:48 and 01:54 are
Research Funds for the Central Universities Project No. 2015CDJXY.

437
436

Authorized licensed use limited to: Fernuniversitaet Hagen. Downloaded on December 13,2023 at 14:05:14 UTC from IEEE Xplore. Restrictions apply.
two check-in in different time, they are in the same time slot location A may be a bar or a pub. Contrary to location A, the
t1. check-ins of location B are mainly in 10:00-14:00 and 16:00-
22:00, which indicates that location B may be a restaurant.
We denote the set of locations by L = {l1, l2, l3 … ln},
where li is a location in Foursquare check-in data. A matrix
can be formed by locations and time slots, which is shown in
Fig. 2. In Fig.2, each row represents how many times a
location is checked-in at different time slots. It should be
noted that the numbers in the matrix represent that the
location is checked-in by all users at that time. Empirically,
the check-in times in different time slots can reflect the
feature of li. For example, an restaurant may be often
checked-in at noontime and evening. If two locations have
similar check-in times in each time-slot, the two locations
can be considered as two similar locations.

Fig. 3. The check-in distributions in different time slots of two locations

According to the examples in Fig. 3, though we cannot

learn the detailed information of the locations, the check-in
times in different time slots can be used to calculate the
similarity of two locations. Given a location li, the check-in
times in different time slots can be considered as a vector,
Fig. 2. Location ¯ Time slot Matrix
denoted by Chk(li). Here, Chk(li) = { Norm(li(t1)),
Norm(li(t2)) … Norm(li(t12))}. Therefore, the similarity
C. Normalization of Location × Time slot Matrix between two locations li and lj, can be determined by
Since the check-in times of different locations are quite calculate the similarity between Chk(li) and Chk(lj).
different, the check-in times should be normalized. For a In this paper, we adopt the most widely used cosine
location li, the check-in times at time slot tj are denoted as method to calculate the similarity between two locations as
li(tj). li(tj) can be normalized as equation (1): in equation (2):
li (t j )
Norm(li (t j )) = 12
(1)
¦ Chk (l ) × Chk (l )
¦ l (t ) sim(li , l j ) = (2)
i j

j =1
i j
¦ Chk (l ) × ¦ Chk (l )
i
2
j
2

The denominator of the normalization means the total

check-in times of location li at all the 12 time slots. In this
way, the Norm(li(tj)) will be limited in [0, 1). Here, sim(li, lj) is in [0, 1]. If sim(li, lj) is close to 1, then
li and lj are similar, otherwise, li and lj are different.

B. The feature locations of user

IV. THE POI RECOMMENDATION METHOD USING
LOCATION SIMILARITY POI recommendation aims to recommend the interesting
places that user has not visited before. Given a user u, we
A. The similarity between two locations denote the set of locations that user u has visited by Lu={lu1,
lu2, … lun}, and denote the set of locations that user u has not
As mentioned in section III, the check-in times in visited by L’u={ l’u1, l’u2, …l’um}. Obviously, Lu and Lu’ are
different time slots can reflect the feature of a location. Fig. 3 the subset of L, and Lu + L’u = L. Therefore, for user u, the
shows two example of the distributions of check-in times in POI recommendation is to generate a location list from Lu’,
different time slots. The distributions of location A and B and recommend the list to user u.
have quite different. For location A, the check-ins are mainly
in 18:00-4:00. Although, the Foursquare check-in data does For user u, the total check-in times of each location lui is
not show what the locations are, however, we can infer that counted, denoted Cnt(lui). All the Cnt(lui) of set Lu are sorted

438
437

Authorized licensed use limited to: Fernuniversitaet Hagen. Downloaded on December 13,2023 at 14:05:14 UTC from IEEE Xplore. Restrictions apply.
from high to low. The top k of Cnt(lui) are chosen as the U
feature locations of user u, denoted flui ( 1  i  k ) ¦ precision u

C. The POI Recommendation method using location

precisionall = u
(6)
U
similarity
For each user u, we calculate the score of each location |U |
l’ui as equation (3):
¦ recall u
recallall = u
(7)
U
Score(u, l ) = Max( sim( fluj , l ))(1 ≤ j ≤ k )
'
ui
'
ui (3)
_U_ represents the number of users who are both in
training set and test set.
where Max() represents the maximal value of similarity
between the feature locations and the unvisited locations. For C. Experimental results
example, for an unvisited location l’ui, the similarity between We set  = {10, 20, 30, 40} as the length of
l’ui and all feature locations of user u are calculated. The recommendation list respectively and set the number of
maximal value of these similarity is chosen as the score of feature locations k = {1, 2, 3} respectively. The results on
l’ui. After the score of each location l’ui is calculated, the precision and recall are shown in Fig. 4.
Score(u, l’ui) of all unvisited locations are sorted from high to
low. The top  of l’ui are generated as the POI
recommendation.

V. EXPERIMENTS

A. Dataset
In our experiments, we set a boundary and extracted
65118 check-ins from the Foursquare dataset. The dataset (a) Precision
contains 2307 users and 5418 locations. In order to alleviate
the sparsity, we removed users who visit place fewer than
three, and randomly sampled 54265 check-in data as a
training set, leaving the remaining 10853 check-in data as a
test set.

B. Evaluation Criteria.
In this experiment, we use the precision and recall to
evaluate the performance of our method. The precision and
recall in the recommended list for a specific user u are
defined as in (4) and (5):

Correctu (b)Recall
precisionu = (4)
α Fig. 4. The precision and recall of the experiment result

Correctu According to the experiment results, the following

recallu = (5) discussions can be made:
Lu
(1) The number of feature locations k influences the results
where  represents the length of recommendation list for user very little. Through our observation, the locations that
u and |Lu| represents the number of user u’s unvisited places, have the high rank always have great scores, no matter
and Correctu represents the correct number of the how many feature locations are chosen. Even if some
recommended list locations of user u. The overall precision locations may have greater scores while the feature
and recall are also defined as in (6) and (7). locations increase, these locations are few. Therefore,
when k = {1, 2, 3}, the change of the results is not
significant.

439
438

Authorized licensed use limited to: Fernuniversitaet Hagen. Downloaded on December 13,2023 at 14:05:14 UTC from IEEE Xplore. Restrictions apply.
(2) While  increases, both precisionall and recallall increase. International Conference on Big Data Computing Service and
For a user u, |Lu| is constant, when  increases, Correctu Applications (BigDataService). IEEE, 2016: 121-130.
also increases. Therefore, recallu increases while  [5] Baral R, Li T. Maps: A multi aspect personalized poi recommender
system[C]//Proceedings of the 10th ACM Conference on
increases. As for the precisionu, we notice when  Recommender Systems. ACM, 2016: 281-284.
increase from 10 to 40 (increases 300%), the precisionu [6] Yin H, Zhou X, Cui B, et al. Adapting to user interest drift for poi
(if k=1) increases from 0.005 to 0.01 (only increases recommendation[J]. IEEE Transactions on Knowledge and Data
100%). In other words, when ={20, 30, 40}, the Engineering, 2016, 28(10): 2566-2581.6
change of precisionu is not significant. [7] del Carmen Rodríguez-Hernández M, Ilarri S, Lado R T, et al.
Location-Aware Recommendation Systems: Where We Are and
Where We Recommend to Go[C]//LocalRec@ RecSys. 2015: 1-8.
[8] Cheng Z, Shen J. Just-for-me: An adaptive personalization system for
VI. CONCLUSION AND PROSPECT location-aware social music recommendation[C]//Proceedings of
international conference on multimedia retrieval. ACM, 2014: 185.
In this paper, we conducted a preliminary study on the [9] Li H, Ge Y, Zhu H. Point-of-Interest Recommendations: Learning
similarity of locations, and proposed a time slots based POI Potential Check-ins from Friends[C]//Proceedings of the 22th ACM
recommendation algorithm on LBSN. By analyzing users’ SIGKDD international conference on on Knowledge discovery and
history check-in data, we calculated the similarity between data mining. ACM. 2016.
different number of visited frequently locations and unvisited [10] Chen D, Ong C S, Xie L. Learning Points and Routes to Recommend
location to generate different recommendation list. The Trajectories[C]//Proceedings of the 25th ACM International on
Conference on Information and Knowledge Management. ACM,
experimental result shows that the number of feature 2016: 2227-2232.
locations k effects the results very little. While  increases, [11] Zeng J, Li F, Liu H, et al. A Restaurant Recommender System Based
both precision and recall increase. Therefore, the proposed on User Preference and Location in Mobile Environment[C]// Iiai
method is effective. International Congress on Advanced Applied Informatics. 2016:55-
60.
In the future, we will introduce friend relationship, [12] Guan X, Li C T, Guan Y. Enhanced SVD for Collaborative
distance and weather information. Mining users’ interest Filtering[C]//Pacific-Asia Conference on Knowledge Discovery and
from users’ mobile location trajectory can also make more Data Mining. Springer International Publishing, 2016: 503-514.
personalized recommendation. [13] Hosseini-Pozveh M, Nematbakhsh MA, Movahhedinia N. A
multidimensional approach forcontext-aware recommendation in
mobile commerce. Int’l Journal of Computer Science andInformation
REFERENCES Security, 2009, 3(1):86-91
[1] Yuan Q, Cong G, Ma Z, et al. Time-aware point-of-interest [14] Vico DG, Woerndl W, Bader R. A study on proactive delivery of
recommendation[C]// International ACM SIGIR Conference on restaurant recommendations for Android Smartphones. In: Proc. of
Research and Development in Information Retrieval. ACM, the RecSys 2011 Workshop on PeMA 2011. Chicago, 2011.
2013:363-372. [15] Karatzoglou A, Amatriain X, Baltrunas L. Multiverse
[2] Kurashima T, Iwata T, Hoshide T, et al. Geo topic model: joint recommendation: n-dimensional tensor factorization for
modeling of user's activity area and interests for location context-aware collaborative filtering, Proceedings of the fourth ACM
recommendation[C]//Proceedings of the sixth ACM international conference on Recommender systems. ACM, 2010: 79-86
conference on Web search and data mining. ACM, 2013: 375-384. [16] Sandvig J J, Mobasher B, Burke R D. A Survey of Collaborative
[3] Gao H, Tang J, Hu X, et al. Content-Aware Point of Interest Recommendation and the Robustness of Model-Based Algorithms[J].
Recommendation on Location-Based Social Networks[C]//AAAI. Bulletin of the Technical Committee on Data Engineering, 2008,
2015: 1721-1727. 31(2):3-13.
[4] Habayeb M, Soltanifar B, Caglayan B, et al. A Novel Point of Interest [17] Burke R. Hybrid Recommender Systems: Survey and Experiments[J].
(POI) Location Based Recommender System Utilizing User Location User Modeling and User-Adapted Interaction, 2002, 12(4):331-370.
and Web Interactions[C]//2016in IEEE Second

440
439

Authorized licensed use limited to: Fernuniversitaet Hagen. Downloaded on December 13,2023 at 14:05:14 UTC from IEEE Xplore. Restrictions apply.