=Paper=
{{Paper
|id=Vol-2327/UISTDA1
|storemode=property
|title=User-relative Personalized Tour Recommendation
|pdfUrl=https://ceur-ws.org/Vol-2327/IUI19WS-UISTDA-1.pdf
|volume=Vol-2327
|authors=Prarthana Padia,Bhavya Singhal,Kwan Hui Lim
|dblpUrl=https://dblp.org/rec/conf/iui/PadiaSL19
}}
==User-relative Personalized Tour Recommendation==
https://ceur-ws.org/Vol-2327/IUI19WS-UISTDA-1.pdf
User-relative Personalized Tour Recommendation
Prarthana Padia∗ Bhavya Singhal∗ Kwan Hui Lim
School of Computing and Information School of Computing and Information Information Systems Technology and
Systems Systems Design Pillar
The University of Melbourne The University of Melbourne Singapore University of Technology
ppadia@student.unimelb.edu.au singhalb@student.unimelb.edu.au and Design
kwanhui_lim@sutd.edu.sg
ABSTRACT recommended places while considering the available tour
Tour planning and recommendation is an important but te- budget, time and cost.
dious task for tourists visiting unfamiliar cities and places. There is an abundance of information available on Inter-
While there are various personalized tour recommendation net about travel guides and famous places, but they do not
works, they typically adopt a simple measure of user in- consider the user’s personal interests and preferences nor
terests based on the number of times a user has visited a contemplate the trip’s constraints like time and cost. Despite
place. In this paper, we propose an improved personalized the availability of such online information, people may end
tour recommendation system that considers a user’s interest up spending excessive efforts and time to plan their itinerary,
preferences in specific categories, relative to his/her overall and yet end up with an undesired itinerary thus leaving them
interests. Using a Flickr dataset across eight cities, we com- with an unsatisfactory and frustrating experience.
pared our proposed algorithm against various baselines and In recent times, personalized tour recommendation sys-
experimental results show that our algorithm obtained supe- tems have benefited from the advancement in web technolo-
rior performance in terms of user interest and popularity. gies and geo-location services. The large amount of online
available geo-tagged photos facilitate the modelling of user
CCS CONCEPTS interest, preferences and trip constraints while strategizing
• Information systems → Personalization; Recommender itinerary planning. While many works consider user interest,
systems; Location based services; Data mining; Web applica- they adopt a simple measure based on the number of times
tions. a user has visited a place.
KEYWORDS
Tour Recommendations; Trip Planning; Recommendation Contributions
Systems; Personalization
Unlike earlier works that adopt a simplistic definition of user
ACM Reference Format: interest based on visit counts, this paper proposes a tour
Prarthana Padia, Bhavya Singhal, and Kwan Hui Lim. 2019. User- recommendation system that utilizes a novel user-relative
relative Personalized Tour Recommendation. In Joint Proceedings measure of interest preferences build upon the Orienteering
of the ACM IUI 2019 Workshops, Los Angeles, USA, March 20, 2019 , problem.
6 pages. We propose two variation of user-specific interest prefer-
ences. The first approach aims to recommend an itinerary
1 INTRODUCTION with no prior knowledge about the user by taking advan-
Tour planning is an important task for ensuring satisfac- tage of the large collection of geo tagged photos available
tory visits to unfamiliar cities and places. However, visitors online. Based on photo frequencies of each POI by all users,
are faced with the challenge of identifying popular places it determines the popularity of each POI and suggests the
aligned with their personal interests. In addition, there is an itinerary based on most popular POIs. The second approach
added complexity due to the need to schedule visits to all aims to improve upon the targeted personalization level for
each user. The benefits of this approach include having a
∗ Both authors contributed equally to this research.
tour itinerary that is customized for each user and caters to
the user’s categories of interest. For example, in a city full
of many tourist attractions, this approach caters to specific
IUI Workshops’19, March 20, 2019, Los Angeles, USA
category the user is interested in. For example, if a user has
Copyright © 2019 for the individual papers by the papers’ authors. Copying
permitted for private and academic purposes. This volume is published and
shown to prefer outdoors and beaches more than museums
copyrighted by its editors. and multiplexes, then the recommendation system takes that
into account while building the itinerary.
IUI Workshops’19, March 20, 2019, Los Angeles, USA P. Padia et al.
2 RELATED WORK 3 PROBLEM FORMULATION AND ALGORITHMS
Orienteering problem overview Problem Formulation. Similar to many earlier works [18,
Orienteering problem is a routing problem which can be 21], we model our recommendation problem based on a vari-
viewed as a contest with multiple nodes, where each node ant of the Orienteering problem [13, 23, 25]. In this tour
has some specific score. The goal of the contest is to maxi- recommendation problem, our main objective is to recom-
mize the total score which is gathered by visiting different mend a tour itinerary I = (p1, ..., p N ) that maximizes the
nodes. The contest is time constrained, which needs a strate- total profit from visiting the list of POIs p1 to P N , while
gical plan to choose a subset of nodes visit in sequence, to ensuring that the tour itinerary can be completed within a
maximize collected budget within given time and budget [23]. specific time budget B. Given a set of POIs P, we optimize
For a detailed review of Orienteering problem, [13, 23, 25] for:
reviews orienteering problem and its applications, discusses
and compares published approaches and heuristics of Orien- Õ Õ
Max Pathpi ,p j ηIntu (pi ) + (1 − η)Pop(pi ) (1)
teering problem.
pi ∈P p j ∈P
where Pathpi ,p j = 1 if a path between POI pi and p j is se-
lected as part of the itinerary, and Pathpi ,p j = 0 otherwise.
Tour recommendation variants Intu (pi ) represents a user-specific interest score of how inter-
Tour recommendation is a well-studied field that typically esting POI pi is to user u, while Pop(pi ) indicates the general
focus on maximizing user preferences within the given trip popularity of POI pi . In addition, Equation 1 is subjected to
constraints [6, 14, 15, 17, 19]. For example, [1] proposed a the following constraints: (i) starting and ending at specific
POI recommender system based on offline modelling (user POIs; (ii) connectivity of POIs in the itinerary; (iii) complet-
preferences learnt from her location history) and online rec- ing the itinerary within a specific time or distance budget
ommendation (social opinions learnt from location history B.
of ‘local experts’). [3] modelled tour recommendation as an In addition, Equation 1 is subjected to the following con-
instance of the Generalized Maximum coverage problem. straints:
Building on the same, [5] suggested a solution by exploiting Õ Õ
an instance of Traveling Salesman Problem (TSP). Others Pathps ,pi = Pathp j ,pd = 1 (2)
modelled the tour recommendation problem as an instance pi ∈I p j ∈I
of the Orienteering problem [9, 10], and various variations Constraint 2 ensures that the recommended itinerary starts
based on specific POI visit sequences [12] and POI category at a specific POI ps , and ends at another specific POI pd . In
constraints [2]. A unique method of using POIs and route real-life, this starting and destination POIs would correspond
information as features to a machine learning algorithm to POIs near the hotel that a tourist is staying at.
to recommend probable tour routes was proposed by [8].
Various works also considered real life constraints like POI Õ Õ
availability and travelling time uncertainty [29, 30], queu- Pathpi ,pk = Pathpk ,p j ≤ 1 (3)
ing time awareness [16], visit duration and recency [18], pi ,pk ∈I p j ,pk ∈I
pedestrian crowdedness [26], transport costs [11]. Similarly, Constraint 3 ensures that the recommended itinerary ful-
various web and mobile applications have been developed fills two conditions, namely: (i) all selected paths are con-
for tour recommendation purposes [4, 7, 20, 24, 27]. nected as a full itinerary; and (ii) no POIs are visited more
Differences with earlier work. Our proposed work dif- than once.
fers from these earlier works in several aspects. We auto-
matically derive a measure of user-based interest from the ÕÕ
Cost(pi , p j )Pathpi ,p j ≤ B (4)
user’s photo frequency at POIs of a specific category, rel-
pi ∈I p j ∈I
ative to: (i) The average photo frequencies of other users
at that POI; and (ii) The average photo frequency of that Constraint 4 ensures that the recommended itinerary can
user at all POIs. In contrast, these earlier works either use be completed within a specific time or distance budget B.
time-based user interest, frequency-based user interest or Algorithms and Baselines. We developed an Integer
explicitly mentioned user interest preferences. In addition, program to solve the problem defined in Section 3, along
we improve upon the targeted personalization level for each with novel approaches to defining user interest preferences.
user by recommending customized itineraries that cater to The two proposed approaches based on photo frequency are
the user interest categories. as follows:
User-relative Personalized Tour Recommendation IUI Workshops’19, March 20, 2019, Los Angeles, USA
(i) POI based Photo frequency (PF P ): Given a set of travel relative to the average visit frequency for a better personal-
history of all users U , the system determines the pop- ization.
ularity of the POI using average photo frequency at
each POI. Definition 3: Photo Frequency based User Interest.
(ii) User based photo frequency (PFU ): Given a set of travel As described earlier, the category of a POI p is denoted as
history of a user u, the system determines the popular- Catp . Given that C represents the set of all POI categories,
ity of that category of POI for the user using average the interest of a user u in POI category c is determined as
photos taken by that user at that category of the POI. follows:
In a city, there can be multiple categories of places com- (1) PF P : Õ (f phpx )
prising of multiple POIS. Consider m POIs for a particular f ph
Intu (c) = δ (Catpx = C) ∀c ∈ C (7)
city. Let P = {p1, .., pm } be the set of POIs in that city. Each px ∈Sph ph(px )
POI p has a category Catp (e.g., church, park, beach) and where δ (Catpx = C) = 1, if Catpx = C and 0, otherwise.
latitude/longitude coordinates associated with it. The popu-
larity function Pop(p) that indicates the popularity of a POI
p, based on the average frequency of photos clicked at that (2) PFU :
f ph
Õ (f phux )
POI. We now introduce the key notations and definitions Intu (c) = δ (Catp = C) ∀c ∈ C (8)
used in this work. u x ∈Sph ,p ∈P ph(u x )
where δ (Catp = C) = 1, if Catp = C and 0, otherwise.
Definition 1: Travel History.
Briefly, the above equations model the interest of a user in
(1) PF P : Given a user u who has visited n POIs, the travel a particular POI category c based on photo frequency at each
history is modelled as an ordered sequence, Sph = POI of category c, relative to the average photo frequency
((p1, f php1 ), (p2, f php2 )...), with each duplet (px , f phpx ) (of all users and a single user) at the same POI. The reason
comprising the visited POI px , and number of photos is that a user is likely to click more photos of the POI that
at POI px . he/she is interested in.
(2) PFU : Given a POI p visited by n users, the travel record Hence, firstly by calculating how many more (or less)
of POI p is modelled as an ordered sequence, Sph = photos a user has taken, the interest level of this user in POIs
((u 1, f phu1 ), (u 2, f phu2 )...), with each duplet (u x , f phux ) of this category can be determined. Secondly, by calculating
comprising the user u x , and number of photos taken how many more (or less) photos have been taken by all users,
by u x . the overall interest level of all users in POIs of this category
Definition 2: Average POI Photo Frequency. can be determined.
(1) PF P : Given a set of travel history of all users U , the
system determines the popularity of the POI using
average photo frequency at each POI. 4 EXPERIMENT METHODOLOGY
1 Õ Õ Dataset. For our experiments, we utilized the Yahoo! Flickr
ph(p) = (f phpx )δ (Px = P)∀p ∈ P (5) Creative Commons 100M dataset [22, 28], focusing on a
n u ∈U p ∈S
x ph dataset of 814k geo-tagged photos across eight cities in-
where n is the number of photos at POI p by all users cluding Toronto, Osaka, Glasglow, Budapest, Perth, Vienna,
U and δ (px = p) = 1, if px = p and 0, otherwise. Delhi and Edinburgh. As provided in [18], these geotagged
photos were mapped to a list of POIs based on their re-
(2) PFU : Given a set of travel record for all POIs P, we spective Wikipedia entries, i.e., proximity of geo-tagged
determine the preference of user u using average photo photos to Wikipedia entries of POIs based on their lati-
frequency of user u at all POIs P. tude/longitude coordinates. Similarly, the categories of POIs
are based on their respective Wikipedia entries. The dataset
1Õ Õ also comprises information like the geo-location coordinates,
ph(u) = (f phux )δ (Ux = U ) ∀u ∈ U (6)
n p ∈P u ∈S date/timestamp of the photos taken. To ensure accuracy and
x ph
generalizability of results, only photos with the highest geo-
where n is the number of photos taken by user u at all location accuracy have been chosen.
POIs P and δ (u x = u) = 1, if u x = u and 0, otherwise. Evaluation and Metrics. We evaluated our algorithm
In tour recommendation problems, the user interest pref- and the baselines using leave-one-out cross-validation, which
erences are typically derived from POI visit frequency [3, 5, involves evaluating a specific travel sequence of a user while
8, 16]. In contrast, we consider a user’s POI visit frequency using his/her other travel sequences as training data. The
IUI Workshops’19, March 20, 2019, Los Angeles, USA P. Padia et al.
Table 1: Comparison between Time-based User Interest (PT − .5T and PT − 1T ), Photo Frequency-based User Interest with
respect to all users (PT − .5PA and PT − 1PA ) and Photo Frequency-based User Interest with respect to a single user (PT − .5PU
and PT − 1PU ).
Osaka Toronto
Algorithm Popularity Interest Precision Recall F1 Score Algorithm Popularity Interest Precision Recall F1 Score
PT − .5PU 1.107 ± .939 1.551 ± .228 0.652 ± .037 0.739 ± .027 0.685 ± .033 PT − .5PU 2.015 ± .062 1.803 ± .084 0.680 ± .013 0.761 ± .009 0.709 ± .011
PT − 1PU 0.772 ± 0.068 1.576 ± .228 0.581 ± .032 0.661 ± .024 0.608 ± .028 PT − 1PU 1.574 ± .047 1.898 ± .088 0.675 ± .013 0.730 ± .010 0.691 ± .011
PT − .5PA 1.118 ± .093 1.652 ± .235 0.650 ± .037 0.752 ± .025 0.689 ± .032 PT − .5PA 2.022 ± .064 1.863 ± .086 0.679 ± .013 0.763 ± .009 0.710 ± .011
PT − 1PA 0.772 ± .067 1.683 ± .237 0.585 ± .032 0.676 ± .023 0.618 ± .027 PT − 1PA 1.517 ± .047 2.012 ± .089 0.672 ± .013 0.730 ± .010 0.689 ± .011
PT − .5T 1.144 ± .092 1.171 ± .205 0.662 ± .037 0.759 ± .026 0.699 ± .033 PT − .5T 1.960 ± .064 1.223 ± .061 0.706 ± .013 0.779 ± .010 0.732 ± .012
PT − 1T 0.737 ± .067 1.205 ± .211 0.622 ± .032 0.682 ± .025 0.641 ± .029 PT − 1T 1.420 ± .043 1.350 ± .069 0.710 ± .013 0.744 ± .011 0.718 ± .012
Glasgow Edinburgh
Algorithm Popularity Interest Precision Recall F1 Score Algorithm Popularity Interest Precision Recall F1 Score
PT − .5PU 1.552 ± .126 1.181 ± .216 0.744 ± .030 0.806 ± .021 0.766 ± .026 PT − .5PU 1.961 ± .052 2.499 ± .115 0.599 ± .012 0.729 ± .008 0.645 ± .010
PT − 1PU 1.113 ± .093 1.199 ± .205 0.708 ± .030 0.730 ± .024 0.707 ± .027 PT − 1PU 1.482 ± .050 2.543 ± .123 0.558 ± .012 0.664 ± .009 0.590 ± .010
PT − .5PA 1.591 ± .947 1.077 ± 1.511 0.764 ± .225 0.802 ± .185 0.764 ± .225 PT − .5PA 1.975 ± .042 2.525 ± .092 0.594 ± .010 0.726 ± .007 0.641 ± .009
PT − 1PA 1.075 ± .087 1.151 ± .192 0.708 ± .029 0.728 ± .024 0.708 ± .026 PT − 1PA 1.461 ± .049 2.582 ± .125 0.554 ± .012 0.662 ± .009 0.587 ± .010
PT − .5T 1.578 ± .125 0.614 ± .106 0.778 ± .028 0.821 ± .020 0.794 ± .025 PT − .5T 2.007 ± .054 1.568 ± .089 0.652 ± .012 0.739 ± .008 0.670 ± .010
PT − 1T 1.001 ± .066 0.676 ± .135 0.736 ± .030 0.739 ± .026 0.727 ± .027 PT − 1T 1.297 ± .049 1.660 ± .103 0.594 ± .011 0.660 ± .010 0.611 ± .010
Perth Delhi
Algorithm Popularity Interest Precision Recall F1 Score Algorithm Popularity Interest Precision Recall F1 Score
PT − .5PU 1.847 ± .190 1.680 ± .263 0.693 ± .047 0.772 ± .037 0.722 ± .042 PT − .5PU 1.647 ± .166 1.294 ± .316 0.731 ± .047 0.804 ± .034 0.757 ± .041
PT − 1PU 1.289 ± .166 1.765 ± .332 0.626 ± .040 0.694 ± .036 0.650 ± .037 PT − 1PU 1.139 ± .145 1.422 ± .352 0.610 ± .042 0.671 ± .036 0.630 ± .038
PT − .5PA 1.786 ± .195 2.025 ± .302 0.680 ± .051 0.780 ± .037 0.718 ± .045 PT − .5PA 1.559 ± .134 1.275 ± .334 0.727 ± .047 0.793 ± .036 0.750 ± .042
PT − 1PA 1.283 ± .196 1.988 ± .271 0.610 ± .049 0.697 ± .038 0.641 ± .043 PT − 1PA 1.130 ± .109 1.332 ± .346 0.614 ± .038 0.676 ± .030 0.636 ± .034
PT − .5T 1.828 ± .168 1.595 ± .206 0.759 ± .041 0.827 ± .029 0.784 ± .036 PT − .5T 1.610 ± .133 0.954 ± .252 0.746 ± .045 0.807 ± .036 0.769 ± .041
PT − 1T 1.274 ± .170 1.710 ± .272 0.677 ± .047 0.740 ± .038 0.699 ± .042 PT − 1T 1.128 ± .100 1.000 ± .256 0.632 ± .042 0.674 ± .036 0.648 ± .039
Budapest Vienna
Algorithm Popularity Interest Precision Recall F1 Score Algorithm Popularity Interest Precision Recall F1 Score
PT − .5PU 2.871 ± .297 3.254 ± .454 0.520 ± .038 0.662 ± .024 0.568 ± .033 PT − .5PU 1.513 ± .052 2.470 ± .122 0.604 ± .013 0.707 ± .010 0.637 ± .012
PT − 1PU 2.106 ± .293 3.216 ± .486 0.503 ± .042 0.616 ± .031 0.530 ± .037 PT − 1PU 1.175 ± .049 2.561 ± .134 0.562 ± .012 0.652 ± .010 0.588 ± .011
PT − .5PA 2.697 ± .308 3.357 ± .484 0.524 ± .037 0.653 ± .025 0.568 ± .032 PT − .5PA 1.573 ± .052 2.460 ± .126 0.614 ± .013 0.710 ± .010 0.644 ± .012
PT − 1PA 2.082 ± .286 3.402 ± .490 0.499 ± .043 0.614 ± .031 0.536 ± .037 PT − 1PA 1.168 ± .049 2.608 ± .134 0.562 ± .012 0.652 ± .009 0.587 ± .011
PT − .5T 2.791 ± .293 1.850 ± .309 0.551 ± .042 0.663 ± .028 0.589 ± .036 PT − .5T 1.577 ± .054 1.576 ± .103 0.629 ± .013 0.713 ± .010 0.656 ± .011
PT − 1T 1.806 ± .226 2.019 ± .334 0.558 ± .037 0.624 ± .029 0.580 ± .032 PT − 1T 1.022 ± .042 1.690 ± .111 0.596 ± .012 0.651 ± .010 0.609 ± .011
starting/ending POI and travel duration are set to that of the (5) Tour F 1 −score: TF1 (I ). The harmonic mean of both
specific travel sequence being evaluated, which is used as the recall and precision of a recommended tour itinerary
a representation of a person’s real-life visit. The following I.
evaluation metrics were used:
(1) Tour Popularity: Tpop (I ). The total popularity of all
POIs in itinerary I. 5 RESULTS AND DISCUSSION
(2) Tour Interest: TIunt (I ). The total interest of all POIs Table 1 show the comparison between Time-based User In-
in itinerary I to user u. terest (PT − .5T and PT − 1T ), Photo Frequency-based User
(3) Tour Precision: Tp (I ). The proportion of POIs recom- Interest with respect to all users (PT − .5PA and PT − 1PA )
mended in itinerary I, which matched user’s real-life and Photo Frequency-based User Interest with respect to
travel sequence. single user (PT − .5PU and PT − 1PU ). Overall results show
(4) Tour Recall: Tr (I ). The proportion of POIs in a user’s that our Photo Frequency-based algorithms outperform the
actual travel sequence that were also recommended in baselines in terms of popularity and interest, while offering
itinerary I. comparable performance in terms of other metrics.
User-relative Personalized Tour Recommendation IUI Workshops’19, March 20, 2019, Los Angeles, USA
Comparison of Popularity and Interest. In terms of equal stand; and (ii) The proposed algorithms outperform
Interest (Tint ) metric, the proposed algortihms outperform one of the baseline algorithms of time-based user interest in
the baselines in all the cities, with PT − 1PA being the best terms of overall precision, recall and F1-score.
performing algorithm, followed by PT − 1PU . In terms of In this work, our focus was to recommend user-relative
Popularity (Tpop ) metric, PT − .5PU performs equally good personalized tour itineraries. Some possible future work
as the baseline PT − .5T . In four cities PT − .5PU performs directions are: (i) Using sentiment awareness on content
the best and in other four PT − .5T . The results show that obtained from location based social network tags to com-
the Photo Frequency-based algorithms reflects on the overall bine sentiment analysis techniques with personalization ap-
popularity and overall interests of the POIs better than the proaches and path planning algorithms for recommending
baselines, for recommending itineraries. itineraries that consider user interest preferences and senti-
Comparison between Time based and Photo-Frequency ments; and (ii) Recommending tour itineraries by considering
based User Interest. With regards to precision (Tp ), recall the public transport arrival and departure time to facilitate
(Tr ) and f1-score (TF1 ) the proposed algorithms outperform realistic tour planning and minimize public transport wait-
one of the baselines (PT − 1T ) in all the cities, standing the ing time. Moreover, real time uncertainty of public transport
second best. In terms of Tp , PT − .5PU performs the second could be modelled to improve the suitability.
best in five cities, followed by PT − .5PA performing the sec-
ond best in one city. In terms of Tr , PT − .5PU performs the ACKNOWLEDGMENTS
second best in four cities and PT − .5PA performs the second This research is partly supported by the Singapore University
best in rest four. Here, PT − .5PU and PT − .5PA perform of Technology and Design under grant SRG-ISTD-2018-140,
equally good. These two proposed algorithms outperform and Defence Science and Technology, Australia. The authors
the baseline (PT − 1T ) with regards to recall. In concern to thank Jeffrey Chan, Aaron Harwood and the anonymous
the F1-score TF1 metric, the proposed algorithm PT − 1PA reviewers for their useful comments on this work.
performs the best on one city, PT − .5PU performs the sec-
ond best in five cities, followed by PT − .5PA performing the REFERENCES
second best in two cities. The results show that the proposed [1] Jie Bao, Yu Zheng, and Mohamed F. Mokbel. 2012. Location-based and
algorithms outperforms the baselines in most cities in terms preference-aware recommendation using sparse geo-social networking
of popularity and interest. The algorithms (PT − .5PU and data. In Proceedings of the 20th International Conference on Advances in
PT − .5PA ) outperform one baseline, performing second best Geographic Information Systems (SIGSPATIAL’12). 199–208.
in terms of precision, recall and f1-score. [2] Paolo Bolzoni, Sven Helmer, Kevin Wellenzohn, Johann Gamper, and
Periklis Andritsos. 2014. Efficient itinerary planning with category
constraints. In Proceedings of the 22nd ACM SIGSPATIAL International
6 CONCLUSION AND FUTURE WORK
Conference on Advances in Geographic Information Systems (SIGSPA-
We modelled our tour recommendation problem as an in- TIAL’14). 203–212.
stance of the Orienteering problem and proposed two al- [3] Igo Brilhante, Jose Antonio Macedo, Franco Maria Nardini, Raffaele
gorithms for recommending personalized tours. We recom- Perego, and Chiara Renso. 2013. Where shall we go today? Planning
touristic tours with TripBuilder. In Proceedings of the 22nd ACM In-
mend suitable POIs using both photo frequency user interest ternational Conference on Information and Knowledge Management
preference and POI popularity. We used geo-tagged photos (CIKM’13). 757–762.
to determine the photo frequency of the user and automat- [4] Igo Brilhante, Jose Antonio Macedo, Franco Maria Nardini, Raffaele
ically derive user interest and POI popularity to train the Perego, and Chiara Renso. 2014. TripBuilder: A Tool for Recommend-
algorithms. Our work improves upon the previous research ing Sightseeing Tours. In Proceedings of the 36th European Conference
on Information Retrieval (ECIR’14). 771–774.
in following ways: (i) We introduce photo frequency based [5] Igo Ramalho Brilhante, Jose Antonio Macedo, Franco Maria Nardini,
user interest derived from the number of photos taken by Raffaele Perego, and Chiara Renso. 2015. On planning sightseeing
the user at a POI of a specific category, unlike earlier works tours with TripBuilder. Information Processing & Management 51, 2
which consider time-based or frequency-based user interest; (2015), 1–15.
and (ii) We improve upon the targeted personalization level [6] Chao Chen, Daqing Zhang, Bin Guo, Xiaojuan Ma, Gang Pan, and Zhao-
hui Wu. 2015. TripPlanner: Personalized Trip Planning Leveraging
for each user by recommending customized itineraries that Heterogeneous Crowdsourced Digital Footprints. IEEE Transactions
cater to the user interest preferences learnt from the user on Intelligent Transportation Systems 16, 3 (2015), 1259–1273.
photo frequency dataset. Using Flickr dataset across eight [7] Dawei Chen, Dongwoo Kim, Lexing Xie, Minjeong Shin, Aditya Kr-
cities, we evaluate our algorithms in terms of precision, re- ishna Menon, Cheng Soon Ong, Iman Avazpour, and John Grundy.
call, F1-score, tour popularity and interest. The results show 2017. PathRec: Visual Analysis of Travel Route Recommendations. In
Proceedings of the Eleventh ACM Conference on Recommender Systems
that: (i) Using photo-frequency based user interest outper- (RecSys’17). 364–365.
form the baselines in all cities in terms of interest. It is at [8] Dawei Chen, Cheng Soon Ong, and Lexing Xie. 2016. Learning Points
par with the baselines in terms of tour popularity by sharing and Routes to Recommend Trajectories. In Proceedings of the 25th ACM
IUI Workshops’19, March 20, 2019, Los Angeles, USA P. Padia et al.
International Conference on Information and Knowledge Management [19] Claudio Lucchese, Raffaele Perego, Fabrizio Silvestri, Hossein Vahabi,
(CIKM’16). 2227–2232. and Rossano Venturini. 2012. How random walks can help tourism. In
[9] Munmun De Choudhury, Moran Feldman, Sihem Amer-Yahia, Nadav Proceedings of the 34th European Conference on Information Retrieval
Golbandi, Ronny Lempel, and Cong Yu. 2010. Automatic construction (ECIR’12). 195–206.
of travel itineraries using social breadcrumbs. In Proceedings of the 21st [20] Ioannis Refanidis, Christos Emmanouilidis, Ilias Sakellariou, Anasta-
ACM Conference on Hypertext and Hypermedia (HT’10). 35–44. sios Alexiadis, Remous-Aris Koutsiamanis, Konstantinos Agnantis, Ai-
[10] Munmun De Choudhury, Moran Feldman, Sihem Amer-Yahia, Nadav milia Tasidou, Fotios Kokkoras, and Pavlos S. Efraimidis. 2014. myVis-
Golbandi, Ronny Lempel, and Cong Yu. 2010. Constructing travel itPlanner GR: Personalized Itinerary Planning System for Tourism.
itineraries from tagged geo-temporal breadcrumbs. In Proceedings of In Proceedings of the 8th Hellenic Conference on Artificial Intelligence
the 19th International Conference on World Wide Web (WWW’10). 1083– (SETN’14). 615–629.
1084. [21] Kendall Taylor, Kwan Hui Lim, and Jeffrey Chan. 2018. Travel Itinerary
[11] Cheng-Yao Fu, Min-Chun Hu, Jui-Hsin Lai, Hsuan Wang, and Ja-Ling Recommendations with Must-see Points-of-Interest. In Proceedings of
Wu. 2014. TravelBuddy: Interactive Travel Route Recommendation the 2018 Web Conference Companion (WWW’18). 1198–1205.
with a Visual Scene Interface. In Proceedings of the 20th International [22] Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde,
Conference on Multimedia Modeling (MMM’14). 219–230. Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. 2016. YFCC100M:
[12] Aristides Gionis, Theodoros Lappas, Konstantinos Pelechrinis, and The New Data in Multimedia Research. Commun. ACM 59, 2 (2016),
Evimaria Terzi. 2014. Customized tour recommendations in urban 64–73.
areas. In Proceedings of the 7th ACM International Conference on Web [23] Theodore Tsiligirides. 1984. Heuristic methods applied to Orienteering.
Search and Data Mining (WSDM’14). 313–322. Journal of the Operational Research Society 35, 9 (1984), 797–809.
[13] Aldy Gunawan, Hoong Chuin Lau, and Pieter Vansteenwegen. 2016. [24] Pieter Vansteenwegen, Wouter Souffriau, Greet Vanden Berghe, and
Orienteering Problem: A survey of recent variants, solution approaches Dirk Van Oudheusden. 2011. The city trip planner: An expert system
and applications. European Journal of Operational Research 255, 2 (2016), for tourists. Expert Systems with Applications 38, 6 (2011), 6540–6546.
315–332. [25] Pieter Vansteenwegen, Wouter Souffriau, and Dirk Van Oudheusden.
[14] Takeshi Kurashima, Tomoharu Iwata, Go Irie, and Ko Fujimura. 2010. 2011. The Orienteering problem: A survey. European Journal of Oper-
Travel route recommendation using geotags in photo sharing sites. In ational Research 209, 1 (2011), 1–10.
Proceedings of the 19th ACM International Conference on Information [26] Xiaoting Wang, Christopher Leckie, Jeffery Chan, Kwan Hui Lim, and
and Knowledge Management (CIKM’10). 579–588. Tharshan Vaithianathan. 2016. Improving Personalized Trip Recom-
[15] Takeshi Kurashima, Tomoharu Iwata, Go Irie, and Ko Fujimura. 2013. mendation to Avoid Crowds Using Pedestrian Sensor Data. In Pro-
Travel route recommendation using geotagged photos. Knowledge and ceedings of the 25th ACM International Conference on Information and
Information Systems 37, 1 (2013), 37–60. Knowledge Management (CIKM’16). 25–34.
[16] Kwan Hui Lim, Jeffrey Chan, Shanika Karunasekera, and Christopher [27] Wolfgang Wörndl and Alexander Hefele. 2016. Generating Paths
Leckie. 2017. Personalized Itinerary Recommendation with Queu- Through Discovered Places-of-Interests for City Trip Planning. In
ing Time Awareness. In Proceedings of the 40th International ACM Information and Communication Technologies in Tourism. Springer
SIGIR Conference on Research and Development in Information Retrieval International Publishing, 441–453.
(SIGIR’17). 325–334. [28] Yahoo! Webscope. 2014. Yahoo! Flickr Cre-
[17] Kwan Hui Lim, Jeffrey Chan, Shanika Karunasekera, and Christopher ative Commons 100M Dataset (YFCC-100M).
Leckie. In Press. Tour Recommendation and Trip Planning using http://webscope.sandbox.yahoo.com/catalog.php?datatype =i&did=67.
Location-based Social Media: A Survey. Knowledge and Information [29] Chenyi Zhang, Hongwei Liang, and Ke Wang. 2016. Trip Recommen-
Systems (In Press). dation Meets Real-World Constraints: POI Availability, Diversity, and
[18] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Traveling Time Uncertainty. ACM Transactions on Information Systems
Karunasekera. 2018. Personalized Trip Recommendation for Tourists 35, 1 (2016), 5.
based on User Interests, Points of Interest Visit Durations and Visit [30] Chenyi Zhang, Hongwei Liang, Ke Wang, and Jianling Sun. 2015. Per-
Recency. Knowledge and Information Systems 54, 2 (2018), 375–406. sonalized Trip Recommendation with POI Availability and Uncertain
Traveling Time. In Proceedings of the 24th ACM International Conference
on Information and Knowledge Management (CIKM’15). 911–920.