=Paper= {{Paper |id=Vol-2435/paper6 |storemode=property |title=TourWithMe: Recommending Peers to Visit Attractions Together |pdfUrl=https://ceur-ws.org/Vol-2435/paper6.pdf |volume=Vol-2435 |authors=Sebastian Vallejos,Marcelo Gabriel Armentano,Luis Berdun |dblpUrl=https://dblp.org/rec/conf/rectour/VallejosAB19 }} ==TourWithMe: Recommending Peers to Visit Attractions Together== https://ceur-ws.org/Vol-2435/paper6.pdf

RecTour 2019, September 19th, 2019, Copenhagen, Denmark. 32

TourWithMe: Recommending peers to visit attractions together
Sebastián Vallejos Marcelo G. Armentano Luis Berdun
ISISTAN Research Institute ISISTAN Research Institute ISISTAN Research Institute
CONICET-UNICEN CONICET-UNICEN CONICET-UNICEN
Tandil, Buenos Aires, Argentina Tandil, Buenos Aires, Argentina Tandil, Buenos Aires, Argentina
sebastian.vallejos@isistan.unicen.edu.ar marcelo.armentano@isistan.unicen.edu.ar luis.berdun@isistan.unicen.edu.ar
ABSTRACT becomes clear given the existence of many websites 1,2,3 and social
When a user travels alone or in a small group, usually likes to share network groups 4,5,6 dedicated to people who wants to meet other
the experience of visiting different attractions in a larger group. people and form groups for tourism.
This article propose TourWithMe, our first approach to the problem In this context, the popularization of mobile devices brings for-
of recommending peers to visit attractions in a city together. To this ward new challenges and opportunities for the implementation of
aim, TourWithMe automatically learns the user’s interests from personalized applications and location-aware services. Particularly,
previously visited attractions, that are then combined with explicit mobile devices enable to capture the user’s mobility history and
preferences provided by the user to find compatible tourists in the taking advantage of geographic proximity of other users to enhance
city. TourWithMe recommends to the user different groups and, for the user experience [14].
each group, attractions that they would enjoy visiting together. In this article, we present TourWithMe, a recommender sys-
tem in the tourism domain that takes advantage of mobile devices
for recommending travellers to form groups to visit attractions or
CCS CONCEPTS points of interest (POI) together. Our approach considers geolocal-
• Information systems → Recommender systems; Social rec- ization provided by mobile devices in two ways. On the one hand,
ommendation; Crowdsourcing. the approach implicitly learns the user’s interest from the places
he/she visits, the amount of time spent in each place, and the time
spent travelling to those places. In this way, users do not have to
KEYWORDS manually check-in every place they visit or to explicitly provide
group recommender system; tourism; crowdsourcing; user model- their interests, as required by most of the current approaches. On
ing the other hand, the approach finds other tourists in the proximity of
the user and suggests forming a group with those users who have
similar interests. Once a group is formed, the approach suggests to
visit nearby venues that the whole group would enjoy visiting.
1 INTRODUCTION The remainder of this paper is organized as follows. Section 2
Visiting a new city is always a challenging experience. Among the discusses related works about recommenders system for tourism.
set of touristic attractions available in the city, tourists have to select, Section 3 presents the proposed approach for recommending trav-
and usually prioritize, those that are more appealing according to ellers forming groups to visit attractions together. Finally, Section
their interests, available time and budget. In consequence, planning 4 presents conclusions and future works.
a holiday is usually a stressful activity and travellers relay in the use
of different applications that may support their decision-making 2 RELATED WORK
processes. Recommenders System for tourism is a hot topic that has been
Recommender systems for tourism arisen to cope with the infor- addressed in several works in the last years. These works proposed
mation overload to which tourists face when visiting a new city. In approaches to recommend users to visit a nearby POI or even a tour
this regard, recommender systems have focused on different aspects itinerary. To carry out this task, proposed approaches used different
of the domain, such as recommending hotels [1, 25], routes [10, 16], information, such as the user’s current location, information about
restaurants [9], itineraries [7, 15], and attractions [13, 33, 34]. nearby POIs, user preferences and interests, current day and time,
A hot topic in recommender systems research is the recommen- temporal restrictions, etc. The kind of information used and the
dation of items to groups of users, since recommendations need way in which this information is obtained vary depending on the
to satisfy a group of users as a whole, instead of individual users approach.
[5, 6]. In the field of tourism, recommender systems for groups have In [31] and [19], authors asked users to manually provide their
been proposed for users who travel with a predefined group (for interest and preferences. Both approaches recommend a personal-
example, a group of friends or family travelling together) [2, 11]. ized tour itinerary that fits the user’s interests. To carry out this
To the best of our knowledge, none of the existing approaches
considered the proposal of groups to visit different attractions to- 1 https://www.yourtravelmates.com/
2 https://www.workaway.info/
gether. This kind of recommender system might be extremely useful
3 https://www.couchsurfing.com/
for users who visit a destination alone or in a small group (for ex- 4 https://www.facebook.com/groups/altmtl/
ample, with his/her couple) and who want to meet peers to share 5 https://www.facebook.com/groups/1157818554266712/

the experience of touring together. The need of this kind of service 6 https://www.facebook.com/groups/travellinks/

Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
RecTour 2019, September 19th, 2019, Copenhagen, Denmark. 33

task, [31] used a Greedy algorithm while [19] used an evolutionary 3 SYSTEM DESIGN
algorithm. The main disadvantages of these works are that man- Figure 1 shows a high-level diagram of TourWithMe. As shown in
ually introducing interests may be a stressful task for users, and the diagram, the approach consists of three steps. In the first step
they tend to be reluctant to explicitly provide this kind of informa- (A) the approach infers the user’s interests from the geolocation
tion [26]. For this reason, some works in the literature proposed to data of the user. By knowing the POIs visited by the user, the time
automatically infer the user’s interests by analyzing the previously spent in each place, and the time spent travelling to those places it
visited POIs. is possible to estimate the interest of the user in such places. This
To address this task, some works used check-ins made by user step is detailed in Section 3.1. In the second step (B), when a user
in location based social networks (LBSN) [17, 35] and geotagged requires it, the approach proposes forming a group with nearby
photos from social networks [4, 20, 21] in order to reconstruct the users. The approach uses the profile information of each user to
history of visited POIs. In [4, 17, 20], authors proposed approaches form a cohesive group of users with similar interests. In this sense,
that infer the interests of the user for each POI category according there is more chance of finding a POI that is attractive to everyone
to the number of visited POIs belonging to that category. These in the group. This step is detailed in Section 3.2. Finally, in the third
approaches use these interests to generate a ranking of possible step (C), the approach recommends the top-five POIs to the group
POIs to be visited by the user. In [35], authors proposed a similar by considering the interest information of each user in the group.
approach that infers the user’s interest from Jiepang check-ins data. This step is detailed in Section 3.3.
As the user’s interests may change according to the time of day,
this approach also divides the day into six time slots and calculates 3.1 Inferring the user’s interests
the user’s interests for each time slot separately. In [21], authors
This step consists of analyzing the mobility data of the user in
proposed an approach that calculates the duration of each visit by
order to infer his/her preferences. In order to carry out this task,
considering the timestamps of the first and the last photos took in
TourWithMe takes advantage of modern mobile devices. These
the visited POI. The approach uses this information to estimate the
devices are equipped with several sensors that allow estimating the
user interest for a POI category. For example, if the user spends
location of the user. For example, it is possible to estimate the user
more time in museums than the average time spent by other users,
location by knowing the nearby WiFis or by using the GPS of the
the approach infers that the user is interested in museums.
smartphone. By tracking the user location, TourWithMe detects
As some tourists tend to travel in group, recommending POIs
visits to places, also named stay points. A stay point is defined in
to a group of users instead of to a single user is a useful feature in
the literature as a geographic region where the user stayed over
the tourism domain. Some approaches in the literature address this
a time threshold Ts within a distance threshold D s [24, 29, 32].
feature by combining the users’ profiles into a single group profile
In particular, TourWithMe detects a visit when the user stays for
[12, 27]. In this way, approaches designed for recommending POIs
more than 5 minutes within a distance of 50 meters. Each visit is
to a single profile (usually a user profile) can recommend also POIs
represented as a tuple (C,Ti ,Te ), where C is the centroid of the
to a group by taking the group profile as input. There are two main
geographic area where the user stayed, Ti is the start time of the
approaches to combine user profiles: aggregation, when the resul-
visit and Tf is the end time of the visit.
tant group profile is the union of all the group members preferences;
When a visit is detected, TourWithMe identifies the POI visited
and intersection, when the resultant group profile is the intersec-
by the user, if any. To carry out this task, TourWithMe relies on
tion of all the group members preferences. The approach presented
public data extracted from OpenStreetMap7 (OSM). In particular,
by [5] used an hybrid approach for generating recommendations to
TourWithMe uses the Overpass Turbo API8 to query POIs that are
groups of tourists, which combines the demographic information
less than 50 meters away from the visit. If there are no nearby POI,
of users, the ratings of the community and the content-specific
it is considered that the user stayed in some other place (e.g. in a
information about the items. The individual ratings inferred from
store). If there is more than one nearby POI, TourWithMe selects the
the hybrid profile are weighted according to a fixed set of social
POI with the highest score according to Equation 1. This equation
relationships among the members of the group. Finally, the influ-
compares the duration of a visit V of user U and the average time of
enced individual ratings of all members of a group are combined to
visit for a POI P. The average time of visit for P is computed from
estimate a group rating for different items.
previous visits of other users to the same POI. It is important to
To the best of our knowledge, none of the existing approaches
notice that the user can manually modify the visited POI if needed.
considered the proposal of groups to visit different attractions to-
gether. The most similar approach to the one presented in this arti-
cle is the one presented in [22]. In this work, authors proposed an |avдDurationO f V isit(P) − duration(V )|
score(V , P) = 1 − (1)
approach oriented to assisting travel agencies for grouping tourists. avдDurationO f V isit(P)
The approach uses K-means algorithm to cluster a predefined set Once the visit has an associated POI, TourWithMe estimates the
of users into K groups. Each resultant group contains users with interest of the user in that POI. The interest of the user in a POI
similar interest. Then, the approach assigns a tour itinerary from a is a real value between 0 and 1 where 0 means that the user is not
set of predefined tour itineraries to each group of users. However, interested in the POI and 1 corresponds to the maximum interest.
this approach is not useful for a tourist who is alone in an unknown This value is computed according to Equation 2 and considers the
city and wants to meet peers to visit POIs together.
7 https://www.openstreetmap.org/
8 http://overpass-turbo.eu/

Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
RecTour 2019, September 19th, 2019, Copenhagen, Denmark. 34

Figure 1: TourWithMe approach

time that the user spent in the POI (intvisit −t ime ) and the time of returns to his place of lodging. The way to calculate the time ratio
the travel T to that POI (intt r avel −t ime ). for journeys in which the user visit several POIs before returning
his/her place of lodging is detailed in [30].
intvisit −t ime (V , P) + intt r avel −t ime (T , P)
int(T , V , P) = (2) duration(T )
2 intt r avel −t ime (T , V ) = (4)
duration(T ) + duration(V )
To compute the first term of the equation, in [21] authors pro-
By knowing the interest of the user in each POI he/she visited,
posed to compute the ratio between the time spent by the user in
it is possible to estimate his/her interest for each POI category. As
the POI and the average duration of visits to that POI. However, this
POIs are extracted from OSM, they have different pairs of key-value
approach is not useful when a POI has different groups of users who
describing them. For example, {”tourism” : ”museum”}, {”name” :
visit the POI with different average times. For example, a museum
”Le Louvre”}. These pairs of key-value are used to label the POI
can offer 1-hour and 2-hours guided tours. An average of 1.5 hours
with POI categories. For example, "Le Louvre" is categorized as a
is then not representative for a user taking the 1-hour tour nor to
"museum". To calculate the interest of a user for a specified POI
a user taking the 2-hours tour. Furthermore, computing the inter-
category C, TourWithMe calculates the average interest of the
est of a user in a POI in this way doesn’t give a normalized value
user in every POI p belonging to C that he/she previously visited
of the user interest. To overcome the above-mentioned problems,
(Equation 5).
TourWithMe uses the cumulative percentage of duration of visits.
Equation 3 shows how the approach computes intvisit −t ime (V , P) P
p ∈C interest(U , p)
for a visit V to a POI P. For example, if spent 14 minutes in P, and intinf er r ed (U , C) = (5)
|C |
60% of people stayed less than 14 minutes in P, then the interest of
the user in P is 0.6. 3.2 Forming groups
Pdur at ion(V ) For suggesting groups to a user, TourWithMe considers three fac-
d=0
Vd,p tors: geolocalization, user’s preferences and similarity between
intvisit −t ime (V , P) = (3)
Vp users’ interests regarding POIs categories. When the user asks for
where Vd,p is number of visits to POI p with a duration d, and suggestions or when he/she arrives to a new city, TourWithMe first
Vp is the number of visits to POI p. find the set of users S R within a parameter radio R from the user’s
The second term of Equation 2, intt r avel −t ime (T , P), compares current location. If R is not set by the user, TourWithMe considers
the time spent by a user in a POI with respect to the time spent the set of users visiting the same city. S R contains then the set of
travelling to that POI. In [8] authors proposed travel-time ratio candidate users near to the user’s location.
as a way to calculate how much time a user is willing to travel Once the set of candidate users is obtained, it is filtered by the
to perform an activity. In [30], authors found higher travel-time user’s preferences. User’s preferences are a list of restrictions that
ratios for activities in which users are interested, such as sport the user is able to manually fill in his/her profile, and indicate the
and recreation activities. Mapping the conclusions arrived in the system what kind of users are expected to be recommended to the
above-mentioned research to the tourism domain, we can assume target user. These preferences, which are all optional, include:
that if a user travels a long time to visit a given POI, he/she has a • age range: indicates the minimum and maximum age of other
great interest in that POI. Equation 4 details how to calculate this users in the group
ratio for simple journeys in which the user goes to a POI and then • sex: preferred sex of people in the group (male, female, any)

Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
RecTour 2019, September 19th, 2019, Copenhagen, Denmark. 35

• languages: a list of languages that users in the group should 3.3 Recommending POIs
speak Although groups are formed by finding tourists with similar inter-
• country of residence: if the user prefers other users from ests, different users always will have some different interests. To
specific countries address these diverse interests, most approaches in the literature
• children preference: users can indicate whether they prefer build a group interest profile by aggregating or by intersecting
tourists traveling with children or not. the preferences of all group members [11, 13, 27, 28]. From these
Then, if a user established in his/her profile that he/she prefers two options, aggregating preferences is preferable since it allows
other tourists aging between 20 and 30, any candidate whose age introducing serendipity in the recommendations enabling the user
is outside those limits is removed from the set of candidates. The to discover attractions that may not be recommended by a recom-
resulting set S f contains the set of compatible candidates with the mender system for individuals. Serendipitous items are items that
user’s preferences. users would not find by themselves or even look for, but that would
Other kind of preferences included in the user profile are the enjoy consuming. The introduction of serendipity in recommender
following: systems is fundamental to avoid users losing the interest in recom-
• a list of categories of interest: an explicit list of the POI cate- mendations due to a overspecialization of the system in the user’s
gories in which the user manually indicated interest. Cate- already-known interests [18]. This overspecialization, avoids the
gories are taken from the OpenStreetMap Semantic Network recommender system to learn new interests of the user, and enables
[3]. the user to be able to predict by themselves what items would be
• budget: indicates the amount of money the user expects to recommended by the system, reducing in consequence the user’s
spend while visiting attractions. This variable is discretized satisfaction with the recommendations.
in four values (0, $, $$, $$$), indicating free, cheap, moderate, For example, Figure 2 shows a group of three users with their
and expensive POIs, respectively respective interests. By aggregating user interests, the interest of
the resultant group profile in a category Ci is the average interest
The list of categories manually defined by the user and the of the three users in Ci . In the literature, this is known as average
inferred interests (which were obtained as described in Section aggregating strategy [23]. As the interest of user B in C 2 is not
3.1) are combined to define the real interest of a user U in a cate- defined, the interest of the whole group in C 2 is calculated by
gory C (Equation 6). If user U explicitly indicated interest in C (by considering only users A and C. Thus, the resultant group profile
adding it to his/her list of interests), then int(U , C) is the average has a high interest in category C 2 . In this way, if the approach
between 1 and intinf er r ed (U , C). Otherwise, int(U , C) is equals to recommends a POI of C 2 , it will encourage User B to visit a new
intinf er r ed (U , C). kind of POI. Instead, by intersecting user interests, the resultant
group profile will not have any interest value defined for C 2 , since
1+int inf e r r ed (U ,C) not all users of the group have an interest defined in C 2 . Thus, the


 2 , if U is interest in C
approach will encourage users to continue visiting the same kind
int(U , C) =


(6)
 of POIs they already visited before.
int
 inf er r ed (U , C),

otherwise
In the current implementation of TourWithMe, each candidate
user v in S f is ranked by computing the soft cosine similarity with
respect to the target user U (Equation 7). This similarity measure
does not assume that features in the space model are independent
and then introduce the similarity of features into the equation of
the traditional cosine similarity.
PN Figure 2: Aggregation vs. intersection of interests
i, j si j Ui v j
so f t_cosine(U , v) = qP qP (7)
N N TourWithMe builds a group interest profile based on the av-
i, j si j Ui U j i, j si j c i v j
erage interest preference of all group members. Given a group
where Ui is the ith feature for user U , vi is the ith feature for д = u 1 , ..., uk , the group interest in a cagetory c is defined accord-
user v, and si j is the similarity between the ith and the jth features. ing to Equation 8.
The similarity between features i and j, si j , is computed by using
the semantic similarity of OSM tags [3]. The set SC ⊂ S f with the 1 X
int(д, c) = int(u, c) (8)
K most similar users is considered for forming groups in the next |дc | u ∈дc
step. where дc ⊂ д are the members of д for which the interest int(u, c)
When a user U asks for a group recommendation, he/she must is defined.
define a preferred group size Z (where Z < K). Then, from Sc , all Then, the interest of a group д in a given POI p is computed
possible groups of size Z including the target user U are computed, according to Equation 9.
and a cohesion score is assigned to each of them. Cohesion is com-
c ∈Cp int(д, c)
P
puted as the average similarity between each pair of users in the
int(д, p) = (9)
group. Groups are finally sorted by the cohesion score. Cp

where Cp are the categories associated to POI p. REFERENCES
Continuing with the example of Figure 2, by using the average [1] Marie Al-Ghossein, Talel Abdessalem, and Anthony Barré. 2018. Cross-Domain
interest of all the group members not necessary may lead to making Recommendation in the Hotel Sector. In Proceedings of RecTour 2018, Julia Neid-
hardt, Wolfgang Wörndl, Tsvi Kuflik, and Markus Zanker (Eds.). 1–6.
the best recommendation. For example, Figure 2 shows that the [2] Aris Anagnostopoulos, Reem Atassi, Luca Becchetti, Adriano Fazzone, and Fab-
group profile has an interest of 0.67 for C 1 and 0.65 for C 3 . Thus, rizio Silvestri. 2016. Tour recommendation for groups. Data Mining and Knowl-
edge Discovery 31, 5 (sep 2016), 1157–1188. https://doi.org/10.1007/s10618-016-
recommending a POI belonging to C 1 would be preferred than 0477-7
recommending a POI belonging to C 3 . However, the variation of [3] Andrea Ballatore, Michela Bertolotto, and David C. Wilson. 2012. Geographic
interests for this category is very high: User A has an interest of knowledge extraction and semantic similarity in OpenStreetMap. Knowledge and
Information Systems 37, 1 (oct 2012), 61–81. https://doi.org/10.1007/s10115-012-
0.81, while User B has an interest of 0.55. Thus, visiting a POI 0571-0
belonging to C 1 seems to be unfair for User B. Moreover, User A [4] Igo Ramalho Brilhante, Jose Antonio Macedo, Franco Maria Nardini, Raffaele
will want to stay in the POI a longer time, while user B will want to Perego, and Chiara Renso. 2015. On planning sightseeing tours with TripBuilder.
Information Processing & Management 51, 2 (mar 2015), 1–15. https://doi.org/10.
leave before. Instead, when visiting a POI belonging to C 3 , the three 1016/j.ipm.2014.10.003
users will have a similar interest in the POI and there are more [5] Ingrid Christensen, Silvia Schiaffino, and Marcelo Armentano. 2016. Social group
recommendation in the tourism domain. Journal of Intelligent Information Systems
chances that they will agree about how long to be in that place. 47, 2 (mar 2016), 209–231. https://doi.org/10.1007/s10844-016-0400-0
To considering this situation, TourWithMe looks for recommend- [6] Amra Delic, Julia Neidhardt, Thuy Ngoc Nguyen, and Francesco Ricci. 2018.
ing the POI that best fits the group profile at the same time that it An observational user study for group recommender systems in the tourism
domain. Information Technology & Tourism 19, 1-4 (feb 2018), 87–116. https:
reduces the variation of interest among users for the recommended //doi.org/10.1007/s40558-018-0106-y
POI. Equation 10 shows how TourWithMe score a POI p for a group [7] Linus W Dietz and Achim Weimert. 2018. Recommending Crowdsourced Trips
д. All POIs in the user’s neighbourhood are ranked according to on wOndary. In Proceedings of RecTour 2018, Julia Neidhardt, Wolfgang Wörndl,
Tsvi Kuflik, and Markus Zanker (Eds.). 13–17.
this equation, and the top-5 POIs are assigned to each group as [8] Martin Dijst and Velibor Vidakovic. 2000. Travel time ratio: the key factor of
recommendations. spatial reach. Transportation 27, 2 (2000), 179–199. https://doi.org/10.1023/a:
1005293330869
[9] Haoxian Feng and Thomas Tran. 2018. Context-Aware Approach for Restaurant
maxInterest(д, p) − minInterest(д, p) Recommender Systems. In Encyclopedia of Information Science and Technology,
score(д, p) = int(д, p)− (10) Fourth Edition. IGI Global, 1757–1771. https://doi.org/10.4018/978-1-5225-2255-
|д| 3.ch153
where maxInterest(д, p)−minInterest(д, p) is the maximum vari- [10] Ander Garcia, Maria Teresa Linaza, Olatz Arbelaitz, and Pieter Vansteenwegen.
2009. Intelligent Routing System for a Personalised Electronic Tourist Guide. In
ation of interest between the members of group д for POI p. Information and Communication Technologies in Tourism 2009. Springer Vienna,
Along to each recommendation TourWithMe computes the esti- 185–197. https://doi.org/10.1007/978-3-211-93971-0_16
mated time that the group would spend in each POI by using the [11] Inma Garcia, Laura Sebastia, and Eva Onaindia. 2011. On the design of individual
and group recommender systems for tourism. Expert Systems with Applications
cumulative percentage of duration of visits, as detailed in Equation 38, 6 (jun 2011), 7683–7692. https://doi.org/10.1016/j.eswa.2010.12.143
3. In this case, if the group interest in the category of a POI p is, for [12] Inma Garcia, Laura Sebastia, Eva Onaindia, and Cesar Guzman. 2009. A Group
example, 0.6 and the time spent by 60% of the people at the given Recommender System for Tourist Activities. In E-Commerce and Web Technologies.
Springer Berlin Heidelberg, 26–37. https://doi.org/10.1007/978-3-642-03964-5_4
POI is t, we assign t as the estimated time that the group would [13] Damianos Gavalas and Michael Kenteris. 2011. A web-based pervasive recom-
spend at p. mendation system for mobile tourist guides. Personal and Ubiquitous Computing
15, 7 (may 2011), 759–770. https://doi.org/10.1007/s00779-011-0389-x
[14] Damianos Gavalas, Charalampos Konstantopoulos, Konstantinos Mastakas, and
4 CONCLUSIONS AND FUTURE WORK Grammati Pantziou. 2014. Mobile recommender systems in tourism. Journal of
Network and Computer Applications 39 (mar 2014), 319–333. https://doi.org/10.
In this article we presented TourWithMe, a first approach to the 1016/j.jnca.2013.04.006
problem of recommending peers to visit different attractions in a [15] Damianos Gavalas, Charalampos Konstantopoulos, Konstantinos Mastakas, and
group. We believe that our approach might appeal tourists traveling Grammati Pantziou. 2014. A survey on algorithmic approaches for solving
tourist trip design problems. Journal of Heuristics 20, 3 (mar 2014), 291–328.
alone or in small groups to enhance the experience of enjoying the https://doi.org/10.1007/s10732-014-9242-5
attractions offered by a new city. [16] Koji Kawamata and Kenta Oku. 2019. Roadscape-based Route Recommender
TourWithMe is currently in a prototype stage, and is developed as System Using Coarse-to-fine Route Search. Journal of Information Processing 27,
0 (2019), 392–403. https://doi.org/10.2197/ipsjjip.27.392
a native Android application. This application tracks user location [17] K. Kesorn, W. Juraphanthong, and A. Salaiwarakul. 2017. Personalized Attraction
and detect visits when the user stays for more than 5 minutes within Recommendation System for Tourists Through Check-In Data. IEEE Access 5
(2017), 26703–26721. https://doi.org/10.1109/access.2017.2778293
a distance of 50 meters. Then, TourWithMe associates each visit to [18] Denis Kotkov, Jari Veijalainen, and Shuaiqiang Wang. 2016. Challenges of
a POI extracted from OpenStreetMap when possible. In addition, Serendipity in Recommender Systems. In Proceedings of the 12th International
TourWithMe identifies the transport mode of each travel, which in Conference on Web Information Systems and Technologies. SCITEPRESS - Science
and and Technology Publications. https://doi.org/10.5220/0005879802510256
the future may be a useful feature for POI recommendation. For [19] Yohei Kurata and Tatsunori Hara. 2013. CT-Planner4: Toward a More User-
example, if user moves by car, it is possible to recommend more Friendly Interactive Day-Tour Planner. In Information and Communication
distant POIs than if he/she moves on foot. Technologies in Tourism 2014. Springer International Publishing, 73–86. https:
//doi.org/10.1007/978-3-319-03973-2_6
The next step in our research is to evaluate our approach with [20] Kwan Hui Lim. 2015. Recommending Tours and Places-of-Interest based on User
a benchmark dataset. As there is no benchmark dataset available Interests from Geo-tagged Photos. In Proceedings of the 2015 ACM SIGMOD on
PhD Symposium - SIGMOD '15 PhD Symposium. ACM Press. https://doi.org/10.
for POI recommendation for group of users, most works in the 1145/2744680.2744693
literature use datasets with individual ratings and simulate groups. [21] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.
The rating of a simulated group for a POI may be estimated as the 2017. Personalized trip recommendation for tourists based on user interests,
average ratings of the group members. The main challenge after
evaluating the proposed approach with a simulated dataset will
naturally be the validation with real users.

points of interest visit durations and visit recency. Knowledge and Information [29] Xiao Wen Ruan, Shou Chung Lee, and Wen Chih Peng. 2014. Exploring Location-
Systems 54, 2 (may 2017), 375–406. https://doi.org/10.1007/s10115-017-1056-y Related Data on Smart Phones for Activity Inference. In 2014 IEEE 15th Interna-
[22] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and S hanika Karunasekera. tional Conference on Mobile Data Management. IEEE. https://doi.org/10.1109/
2016. Towards next generation touring: Personalized group tours. In Twenty-Sixth mdm.2014.71
International Conference on Automated Planning and Scheduling. 421–430. [30] Yusak O. Susilo and Martin Dijst. 2010. Behavioural decisions of travel-time
[23] Judith Masthoff. 2015. Group Recommender Systems: Aggregation, Satisfaction ratios for work, maintenance and leisure activities in the Netherlands. Trans-
and Group Attributes. In Recommender Systems Handbook. Springer US, 743–776. portation Planning and Technology 33, 1 (feb 2010), 19–34. https://doi.org/10.
https://doi.org/10.1007/978-1-4899-7637-6_22 1080/03081060903429280
[24] Raul Montoliu and Daniel Gatica-Perez. 2010. Discovering human places of [31] Pieter Vansteenwegen, Wouter Souffriau, Greet Vanden Berghe, and Dirk Van
interest from multimodal mobile phone data. In Proceedings of the 9th International Oudheusden. 2011. The City Trip Planner: An expert system for tourists. Expert
Conference on Mobile and Ubiquitous Multimedia - MUM '10. ACM Press. https: Systems with Applications 38, 6 (jun 2011), 6540–6546. https://doi.org/10.1016/j.
//doi.org/10.1145/1899475.1899487 eswa.2010.11.085
[25] Qing Qi, Jian Cao, Yudong Tan, and Quanwu Xiao. 2018. Cross-Domain Recom- [32] Yang Ye, Yu Zheng, Yukun Chen, Jianhua Feng, and Xing Xie. 2009. Mining
mendation Method in Tourism. In 2018 IEEE International Conference on Progress in Individual Life Pattern Based on Location History. In 2009 Tenth International
Informatics and Computing (PIC). IEEE. https://doi.org/10.1109/pic.2018.8706265 Conference on Mobile Data Management: Systems, Services and Middleware. IEEE.
[26] Feng Qiu and Junghoo Cho. 2006. Automatic identification of user interest for https://doi.org/10.1109/mdm.2009.11
personalized search. In Proceedings of the 15th international conference on World [33] Hongzhi Yin, Weiqing Wang, Hao Wang, Ling Chen, and Xiaofang Zhou. 2017.
Wide Web - WWW '06. ACM Press. https://doi.org/10.1145/1135777.1135883 Spatial-Aware Hierarchical Collaborative Deep Learning for POI Recommenda-
[27] Logesh Ravi and Subramaniyaswamy Vairavasundaram. 2016. A Collaborative tion. IEEE Transactions on Knowledge and Data Engineering 29, 11 (nov 2017),
Location Based Travel Recommendation System through Enhanced Rating Pre- 2537–2551. https://doi.org/10.1109/tkde.2017.2741484
diction for the Group of Users. Computational Intelligence and Neuroscience 2016 [34] Hongzhi Yin, Xiaofang Zhou, Bin Cui, Hao Wang, Kai Zheng, and Quoc Viet Hung
(2016), 1–28. https://doi.org/10.1155/2016/1291358 Nguyen. 2016. Adapting to User Interest Drift for POI Recommendation. IEEE
[28] Senjuti Basu Roy, Laks V.S. Lakshmanan, and Rui Liu. 2015. From Group Rec- Transactions on Knowledge and Data Engineering 28, 10 (oct 2016), 2566–2581.
ommendations to Group Formation. In Proceedings of the 2015 ACM SIGMOD https://doi.org/10.1109/tkde.2016.2580511
International Conference on Management of Data - SIGMOD '15. ACM Press. [35] Zhiwen Yu, Huang Xu, Zhe Yang, and Bin Guo. 2016. Personalized Travel Package
https://doi.org/10.1145/2723372.2749448 With Multi-Point-of-Interest Recommendation Based on Crowdsourced User
Footprints. IEEE Transactions on Human-Machine Systems 46, 1 (feb 2016), 151–
158. https://doi.org/10.1109/thms.2015.2446953