Hindawi Publishing Corporation

Computational Intelligence and Neuroscience

Volume 2016, Article ID 1291358, 28 pages
http://dx.doi.org/10.1155/2016/1291358

Research Article
A Collaborative Location Based Travel Recommendation System
through Enhanced Rating Prediction for the Group of Users

Logesh Ravi and Subramaniyaswamy Vairavasundaram

School of Computing, SASTRA University, Thanjavur, Tamil Nadu 613401, India

Correspondence should be addressed to Logesh Ravi; [email protected]

Received 4 November 2015; Revised 18 January 2016; Accepted 31 January 2016

Academic Editor: Silvia Conforto

Copyright © 2016 L. Ravi and S. Vairavasundaram. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Rapid growth of web and its applications has created a colossal importance for recommender systems. Being applied in various
domains, recommender systems were designed to generate suggestions such as items or services based on user interests. Basically,
recommender systems experience many issues which reflects dwindled effectiveness. Integrating powerful data management
techniques to recommender systems can address such issues and the recommendations quality can be increased significantly. Recent
research on recommender systems reveals an idea of utilizing social network data to enhance traditional recommender system
with better prediction and improved accuracy. This paper expresses views on social network data based recommender systems
by considering usage of various recommendation algorithms, functionalities of systems, different types of interfaces, filtering
techniques, and artificial intelligence techniques. After examining the depths of objectives, methodologies, and data sources of
the existing models, the paper helps anyone interested in the development of travel recommendation systems and facilitates future
research direction. We have also proposed a location recommendation system based on social pertinent trust walker (SPTW) and
compared the results with the existing baseline random walk models. Later, we have enhanced the SPTW model for group of users
recommendations. The results obtained from the experiments have been presented.

1. Introduction and accuracy are balanced in the generated recommenda-

tions. Filtering mechanisms play an important role in the
The enormous growth of web and its user base has become recommendation process [7]. The most commonly used
source for large amount of information available online. filtering techniques are collaborative filtering, content-based
This information may be helpful for users, to suggest items filtering, knowledge-based filtering, and social filtering [8].
or services as per their preferences. Recommender system Already many researches had contributed to the development
plays the role of generating suggestions by collecting user of various recommender systems such as movie [9, 10], music
information such as preferences, interests, and locations. [11, 12], books [13], e-commerce [14–16], e-learning [17, 18],
The research on recommender systems gained importance web search [19], and tourism [20].
after the emergence of collaborative filtering [1, 2]. Various The main aim of using personalization techniques is to
researches led to the implementation of recommender sys- generate customized recommendation according to the user
tems to different domains such as e-commerce [3], advertis- preferences and interests [21]. The recommender system has
ing [4], and tourism [5]. Generating suggestions according to an objective to filter unwanted information and to provide
user preferences is a complex task for recommender systems. specific results for the particular user [22]. In the travel
Semantic web technologies help recommender systems to recommender systems [20], proposed model learns the user
resolve those tasks easily [6]. Recommender system uses preferences and generates places of attractions according to
information from many sources to make predictions and to the user interests. This paper focuses on the recommender
suggest an item for a user. Factors such as novelty, stability, systems and their application in tourism. To make this paper
2 Computational Intelligence and Neuroscience

useful to all, including new readers of recommender systems,

it covers topics from evolution to applications along with the Ask for recommendation
Interface
challenges in it. Since more research is required to improve
the effectiveness and efficiency of recommender systems, this
paper will be more useful to the upcoming researchers to

recommendation

Recommender system
Database
develop a user specific recommender system.

Receive
This paper contributes clear review of recommender
systems published in scientific journals and conferences
with a special focus on travel recommender systems. These
systems are analyzed through the recommendation mech- Top n Recommendation
recommendations engine
anism, interface, data source, and functionalities used. The
paper also provides some guidelines to develop efficient, user
specific travel recommender systems. Figure 1: Traditional recommender system.
As a description about the methodology of collection
and organization of articles for the analysis on the travel
recommendation problem, a starting study was performed to of users to make a decision regarding the recommended
focus on the most illustrative subjects and terms in the rec- products or services. By using community-contributed data,
ommender system field. Initially, 105 recommender system such as blogs, social networks, Geographic Positioning
papers were chosen from various journals and conference Systems (GPS) logs, and geotagged photos, recommender
publications, with a higher need for current and frequently systems tend to help the users by generating personalized
referred to articles. Next, we extricated from these papers the recommendations, which will be more useful for the users
most considerable terms. We gave the most accentuation to in their decision making process. Figure 2 describes the
decisive words, less accentuation to titles, and, at last, the conventional work flow model in a recommender system.
minimum accentuation to modified works.
The paper gives a clear view on recommender systems
2.1. Foundations of Recommender Systems. Traditionally, rec-
and explains the various problems in the traditional recom-
ommender systems are based on their building blocks
mender systems. Then, we have explained the development
such as algorithms, filtering methodologies, taxonomies, and
of travel recommender systems along with the techniques
databases. When the recommender systems have only small
and interface types. Later, the location based social network
amount of data for generating suggestions, collaborative
was introduced and its needs and issues were explained in
models face issues with them. Such problem is called cold
detail. As for the part of proposed model, we have explained
start problem and it is described below. Then, the section
in depth about SPTW (social pertinent trust walker) model
describes 𝑘NN algorithm, which is mostly used by collabora-
for category of location recommendations. For the enhance-
tive filtering based recommender systems. The usage of sim-
ment of the proposed location recommendation model, we
ilarity and differences between users’ interests is mostly used
have introduced SPTW based group recommendation model
by many recommendation models. Finally, by comparing the
(SPTW-GRM) for group of users. The proposed models
users or items, different similarity measures were described.
have proved their efficiency and accuracy through evaluation
metrics.
The paper is organized as follows. The next section 2.1.1. Fundamentals. In the recommender systems, process
describes the significance of recommender systems. Section 3 of generating recommendations depends on various factors,
explains briefly about travel recommender systems and such as the following:
Section 4 briefs the use of AI techniques in tourism recom- (i) available user data in the database (such as user
mender systems. Then, Section 5 discusses in brief location information, interests, ratings, locations, and social
based social networks and Section 6 portrays the proposed relationships);
social pertinent trust walker (SPTW) algorithm. Section 7
describes the proposed SPTW based group recommendation (ii) filtering mechanism/algorithm used (like, content-
model (SPTW-GRM) and Section 8 illustrates evaluation of based, hybrid, collaborative, etc.);
the proposed model and discussion on the results. Finally, the (iii) techniques used to enhance the results (such as
paper concludes with the analysis of surveyed systems and Bayesian networks, singular value decomposition,
Section 9 also indicates the new areas to be focused on in the and fuzzy models);
area of travel recommender system in future. (iv) sparsity level and scalability of database;
(v) system performance (such as memory and time con-
2. Significance of Recommender Systems sumption);
Recommender systems (RSs) were generally defined as expert (vi) considered objectives of the system (such as top
systems which are used to recommend products or services recommendations and predictions);
to the users. Figure 1 portrays the working of a traditional (vii) quality and its metrics used for the result and analysis
recommender system. Various factors influence the interests (such as precision, recall, 𝐹-measure, and novelty).
Computational Intelligence and Neuroscience 3

Top n Cross validation

precision; recall;
recommendations ROC rank metrics

Cross validation
Predictions MAE; RMSE;
coverage

Model based Hybrid U2U/I2I

Genetic User to user Item to item
algorithms Fuzzy kNN kNN

Neural Bayesian
networks networks
Memory based

LSI SVD Similarity Aggregation

measures approaches

Hybrid filtering

Content-based filtering Demographic filtering Collaborative filtering

Nonpublic Public databases

commercial LastFM; Netflix
databases Movielens

Figure 2: Conventional work flow model in a recommender system.

Public databases are used in the research of recommender 2.1.3. The 𝑘 Nearest Neighbors Recommendation Algorithm. In
systems to develop new methods, techniques, and algorithms. most of the recommendation processes, using collaborative
Delicious and last.fm are the most popular databases used in filtering, the reference algorithm used is 𝑘 Nearest Neighbors
the development of recommender systems. (𝑘NN) recommendation algorithm. 𝑘NN recommendation
algorithm is simple and reasonably produces accurate results.
2.1.2. Cold Start Problem. The problem of generating nonreli- The drawback of the 𝑘NN recommendation algorithm is
able recommendations due to lack of initial ratings is known low scalability [29] and it is vulnerable to the high level
as cold start problem [22, 23]. New user, new item, and new of sparsity [29, 30] in recommender systems databases. The
community are the three types of cold start problems. During 𝑘NN algorithm focuses on similarity measures and generally
recommender systems’ operations, new user problem [24, 25] similarity computation is carried out between user to user
is a great difficulty in producing personalized recommen- [31], item to item [32, 33], and user to item [34]. Using
dations. Since there are no user ratings provided by these similarity measures, similar users are assigned as neighbors to
new users, memory-based content filtering cannot help in the user and items recommendation is predicted for the user.
the recommendations. New users may reject unreliable, non- Then, from the top-𝑛 recommendations, 𝑛 items are chosen
personalized recommendations and the recommendation to satisfy the particular active user.
services too. Adding additional information to the new user
database, such as preferences, tackles the new user problem. 2.1.4. Similarity Measures. The similarity between users or
Similarly, new item problem [26, 27] arises due to addition items can be determined by similarity metric or similarity
of new items in the recommender systems. Since there is no measure. The most commonly used traditional metrics are
initial rating for these new items added to the recommender cosine (COS), adjusted cosine (ACOS), Pearson correlation
systems, it gets unnoticed by most of the users and large (CORR), constrained correlation (CCOR), Euclidean (EUC),
group of users may be unaware of such items. Developing a and Mean Squared Differences (MSD) [22, 35]. There are also
set of motivational users to rate the new items will help in few new metrics used, such as Jaccard Mean of the Squared
solving new items problem. New community problem [28] Differences (JMSD) [36] calculated by using nonnumerical
occurs during the initialization of recommender systems due information, Singularity (SING) [37] premeditated by utiliz-
to insufficient ratings. Collaborative filtering based recom- ing the information in the user votes, GEN [31] used as a
mendations and encouraging users to rate items can easily similarity measure in the recommender system which uses
solve the new community problem. the genetic algorithms, NCS [37] used as a similarity measure
4 Computational Intelligence and Neuroscience

allows the user to explore maps by zoom, select, and drag-

Hybrid drop options. Web based systems are thus suitable for the
(web + mobile)
planning stages of the tour. The design of these systems limits
the usage to computer systems, the unavailability of which is a
Type of UI

Mobile
concern that a tourist has to keep in mind. The availability of
large desktop screen to provide easily retrievable information
is a key advantage of these systems. We now provide exploited
features of this type of recommender systems. In Venkataiaha
Web
et al. [42], a comparative study has been made between the so-
called discrete and continuous systems. In discrete systems,
0 10 20 30 40 50 the screen is utilized to a high extent providing the users
Distribution (%) with the needed information, whereas the continuous system
combines the associated media, text, photographs, and video
Figure 3: Statistics of interfaces used in the existing systems.
contents into a single video clip, thereby reducing the efforts
required by the user to understand the contents provided.
in the recommender system which uses neural learning, The proposed work of Lee et al. [43] is the first approach
TRUST [38] which utilizes the reputation of users’ ratings that incorporated Google Maps Services for the web interface
on items, and UERROR [39] which predicts the initial actual allowing the plotting of paths on map, guiding the user
ratings of the user and determines the predictions errors too. through the personalized route to the selected locations and
food places at the Tainan City. City Trip Planner [44], e-
Tourism [45], and Otium [46] are some of the web based
3. Travel Recommender Systems systems in which a map is marked with the scheduled
locations to be visited for a single day. EnoSigTur [47] also
The large amount of user information available is exploited
uses the same approach in addition to which the user’s pref-
by the travel recommender systems to provide suggestions to
erences and sociodemographic information is obtained, after
the user in the effective manner [40]. The tourism/travel rec-
which the recommendations are made.
ommender system employs artificial intelligence techniques
An Avatar-based interactive approach that allows the
to generate personalized recommendations to the user. This
users to provide their requirements has been implemented
section depicts the applications of the recommender systems
by the VIBE virtual spa advisor [48]. This allows dynamic
in the field of e-Tourism.
addition of new attributes to the catalogue which in turn auto-
mates the changes in recommendation, preference elicitation,
3.1. Interface. This section presents an overview of different and also the web interface process. It also includes section
interfaces prevalent in recommender systems today. Inter- where domain experts can manipulate the conversational and
faces are web oriented or mobile device based; some are recommendation procedures.
designed with compatibility of both web and mobile device An ontology based recommender system by Wang et al.
in mind. Classification of recommenders is based on two [49] uses semantic web technologies coupled with Web 2.0
major categories (namely, web based and mobile based). services which integrates the different information from the
Figure 3 describes the statistics of interface used in the user during the travel. It is an Ajax web based application
existing systems in percentage. Users were found to give developed on Ruby on Rails which includes other third-party
high preference to web based interface since it provides ease services such as Yahoo weather, Google Map, and Wiki Travel.
of use from any computer without the need of download,
installation, and configuration procedures. But due to the 3.1.2. Mobile Recommendations. Mobile based recommenda-
rapid increase in the number of mobile devices, especially tions system is increasing in the recent years due to the avail-
Smartphones, most web based recommender systems also ability of mobile devices that support Internet facility and the
have a mobile device based counterpart. Smartphones. Mobile recommender systems are designed in
Some systems are desktop specific and they do not such a way that only the relevant and essential information
support web oriented or mobile device based interfaces (e.g., is provided to the user since the Internet connection would
[41]). It has to be noted that these systems are easier to be slow and also the amount of information that can be
develop and implement. But they are not popular among the effectively displayed is lesser than that of a standard web page.
tourists, since downloading, installing, and configuring are It has to be noted that, apart from the modern Smartphones
preliminary parts of their usage. Tourists tend to prefer very which provide easier touch based interaction, in the mobile
simplistic approach to get their recommendations. based systems it is difficult to execute actions such as scrolling
compared to standard web page. The advantage of the mobile
3.1.1. Web Based Recommenders. Web based interface is the based systems is that it can be used in any place where
most prevalent among the interfaces used for e-Tourism Internet connection is available; this can be during the travel
recommender systems. It is highly user friendly with various which allows acquiring of online information that can be
interaction options for the users to get information from used to further enhance the recommendations. The GPS
maps, images, and videos. The possibility of mouse usage which is a part of most of the mobile devices is also used
Computational Intelligence and Neuroscience 5

to locate the user which in turn helps in providing relevant usable in both the iPhone and the iPad devices. Another iOS
recommendations. platform based system is REJA [58].
Yu and Ping Chang [50] report a system that can be
considered the first to implement mobile system based 3.2. Functionalities. We describe the general function fea-
approach which was designed for PDAs. Recommendations tures of the reviewed tourism recommender systems for
based on the user-location and planning suggestions for tour the approaches in four groups which are based on the
were the services provided by the system which took into recommendation of suggestions and the content of tourist
account user’s location, time of access, and preferences. The package provided, suggestion of attractions of a particular
different services provided by the system, the user preferences
location, design of long trips with schedules, and social
setting, the tour plan recommendations, and the usage of
media capabilities. In the subsections that follow we com-
Google Maps can be seen.
ment further about these capabilities with relevant examples.
MTRS [51] is also a PDA based approach which con-
siders the problem of Internet connectivity. The Internet Table 1 summarizes the comparison of travel recommender
connectivity is a problem for the tourist either because of system based on its interface and functionalities.
the rural area or because of the high cost during roaming.
Proximity detection along with a method to update the 3.2.1. Travel Destination and Tourist Packs. User’s preference
content with minimal cost was proposed which use wireless is taken into consideration in some of the systems to provide
sensor networks that would be implemented in small or the recommendation so that it suits the user. PersonalTour
medium scale range. The infrastructure would allow placing [59], Itchy Feet [60], and MyTravelPal [61] are of this type.
higher weight to the ratings of users from fixed connection PersonalTour is a recommender system that is used by travel
facilities compared to users who are away from the visited agencies in order to find suitable travel packages in accor-
locations and rate using Internet. dance with the customer preference. The recommendation
MapMobyRek [52] is another mobile based system that is made in the form of a list of suggestions. Then the rating
takes advantages of the interface by using maps and lists of each travel service from each item can be done by the
to provide the recommendations effectively. Comparison of customer.
places and items based on their characteristics was facilitated Itchy Feet allows the purchase of services that would
by side-by-side display which helped us to decide among book trips, assistance, and other services along with the
recommendations. Another product that serves as a tourist recommendation of the locations. It uses both the internal
guide by describing the recommended locations when the database and external data sources when a user makes a
user is near them is GeOasis [53]. It utilizes the device GPS search request. Its interface allows the user to select from the
to locate the tourist and the travelling speed so that the result items shown such as a list of flights or hotels.
estimation of time can be done to prepare the explanations.
The user is provided with two-way interaction option, using MyTravelPal [61], in accordance with the affinity to user
voice recognition or the tactile interface. areas of interest, is recommended first graphically. When an
Nowadays, the newest technologies that include Android area is chosen, further recommendations of tourist spots and
and iPhone platforms are targeted for the development of services are listed based on the preferences of the user.
mobile tourism recommenders as there is a continuous
increase of its user base. We now show some of the popular 3.2.2. Ranked List of Suggested Attractions. Recommender
systems that make use of these platforms. MoreTourism [54] systems for tourism generally provide suggestions only after
is system that is based on Android which makes use of videos, acquiring the information such as the destination and cost
images, mashups, geolocation, and other available features beforehand from the user. This leads to listing out many
to aid the user in getting information. The EnoSigTur [47] attractions, temporal events, and other places of interest.
is also a system that uses the Android platform for place Hence, these systems are more complex as the system clas-
recommendation, route aiding for trips, and description of sifies and ranks relevant suggestions from a huge database
place of interest. of available information. The suggested list of attractions
The user is able to provide the push information in helps the user to spot places of interests in an efficient
accordance with their context in the LiveCities recommender manner and supports him in discovering more about the
system [55] which makes use of the notification service of the locations. Static database is commonly used for storing the
Android platform. The push information may be text, video, elements for recommendation. Some systems make use of the
audio, or HTML. STS system [56] allows the user to provide web by extracting information automatically so that updated
accurate information of interests, opinions, and descriptions recommendation is ensured (e.g., Otium [46]).
of the visited places and turns it into a powerful application Some recommender systems match the preferences of
based on the Android platform. This is achieved through a the user, check the past travel history for locations, and also
user friendly and intuitive design. compare the positively reviewed locations of other users to
The GUIDEME [57] is a recent system which can be noted provide a suggestion list. This is achieved by the usage of
for its design and implementation, since it has a compatible mechanisms to compare various preferences and similarities
app for both phones and tablet devices. It is designed for between various user profile and streaming data. Contextual
the iOS platform with adaptive features that would allow factors such as user’s current location may also be considered
adjusting with respect to the screen sizes thereby making it to select the recommendations [58]. Justification capability
 6

Table 1: Comparison of interface and functionality of travel recommender systems.

Interface Functionality
Articles Context aware system
Web Mobile Hybrid Attractions suggestion Destination recommendation/tourist package Trip planner Social features
Jannach et al. [48] ✓ ✓
Mı́nguez et al. [62] ✓ ✓
Ricci et al. [52] ✓ ✓
Sebastia et al. [63] ✓ ✓ ✓
Vansteenwegen et al. [44] ✓ ✓ ✓ ✓ ✓ ✓ ✓
Borràs et al. [64] ✓ ✓
Garcia et al. [65] ✓ ✓ ✓ ✓
Gavalas and Kenteris [51] ✓ ✓ ✓ ✓ ✓
Lorenzi et al. [59] ✓ ✓ ✓ ✓
Luberg et al. [66] ✓ ✓ ✓
Martin et al. [55] ✓ ✓ ✓
Montejo-Ráez et al. [46] ✓ ✓ ✓
Rey-López et al. [54] ✓ ✓ ✓ ✓ ✓ ✓
Wang et al. [49] ✓ ✓ ✓
Batet et al. [67] ✓ ✓ ✓ ✓
Borràs et al. [68] ✓ ✓ ✓ ✓ ✓ ✓
Koceski and Petrevska [61] ✓ ✓ ✓
Martı́nez-Santiago et al. [53] ✓ ✓ ✓
Noguera et al. [58] ✓ ✓
Braunhofer et al. [56] ✓ ✓ ✓
Garcia et al. [69] ✓ ✓ ✓
Gyorodi et al. [70] ✓ ✓
Kurata and Hara [71] ✓ ✓ ✓
Lucas et al. [72] ✓ ✓ ✓
Meehan et al. [73] ✓ ✓ ✓ ✓
Ruotsalo et al. [74] ✓ ✓ ✓ ✓ ✓
Savir et al. [75] ✓ ✓ ✓
Umanets et al. [57] ✓ ✓ ✓ ✓ ✓
Yang and Hwang [76] ✓ ✓ ✓ ✓
Computational Intelligence and Neuroscience
Computational Intelligence and Neuroscience 7

for the list of suggestions is also provided by some of the Retrieval of the complete schedule along with the routes is
systems (e.g., [48]). possible by the user once it is completed. Retrieval methods
Automated detection of user’s indoors or outdoors pres- differ among systems, such as EnoSigTur [47] that supports
ence is made in a more complex system SMARTMUSEUM the download in PDF format that contains a map which is
[74], which utilizes the user-location information for this georeferenced along with explanations in detail. City Trip
purpose. When the user is outdoors, the conventional map Planner [44] and Otium [46] support the downloading of the
representation is provided, whereas when the user is indoors, route map and details to a mobile phone.
suitable listing of objects based on the preferences of the user
is given. 3.2.4. Social Aspects. Social functionalities have been focused
on by some of the projects (such as [44, 57, 69, 79]), thus
3.2.3. Planning a Route. Apart from providing the list of spots allowing the user to interact and share material such as
and locations that is relevant to the user preferences, there are evaluation, comments, or pictures with other tourists. This
some systems that guide the tourist in preparing a route plan aspect promotes the usage of recommender when a user is in
along several places. a particular location or activity. Some systems such as Itchy
CT Planner [41, 71] refines the tour plans offered based Feet [60] and MoreTourism [54] allow the user to organize
on the preferences and requests of the users while selecting events or activities with similar tourists apart from interacting
the plans. The user’s choice of factors such as walking speed and commenting. e-Tourism [65] can be made to consider the
and duration is considered to plan the route. A radar chart preferences of an entire group of visitors and suggest activities
and a cartoon character are parts of the interface that helps to or route plans.
navigate providing better interactivity. iTravel [76] allows peer-to-peer communications to share
In some of the systems, the users can build up or manip- ratings and reviews of attractions. The navigation map shows
ulate an initial plan recommended which would include the positions of users nearby along with the route and
activities and locations. The user can reorder the plan and attractions. The system shows the green pins and blue pins
route, add more activities, and schedule them according which reflect the suggested attractions and nearby users,
to their choice. The initial plan is designed by considering respectively.
the timing of various attractions, distances to be covered
Alchemy API is used by the VISIT system [73] in order to
in between, and the expected visit duration. Examples of
apply sentiment analysis methods so that Twitter and Face-
this type of recommender system include EnoSigTur [47],
book updates of a given attraction are analyzed to determine
City Trip Planner [62], Smart City [66], Otium [46], and
whether the users are providing positive or negative emo-
e-Tourism [45]. Vansteenwegen and Souffriau [77] discuss
tional comments regarding it. This information is conveyed to
in detail the functionalities of such trip planners. SAMAP
the users by the display of green and red colors in the interface
[78] and PaTac [79] are some of the advanced systems that
indicating the most loved locations of the day by the users and
generate recommendations, by calculating the possibilities of
those that are not.
activities with respect to different transporting facilities; that
The three main objectives for the usage of social informa-
is, mode of transport, car, bike, walk, or public transport, is
tion are (i) to enhance the prediction in terms of accuracy,
taken into account.
(ii) to develop or design new hybrid recommender systems,
Geographical Information Systems (GIS) are incorpo-
and (iii) to determine the relating features between various
rated by some of these systems for the management of
processes and the social information.
geographical data that has to be associated with the recom-
When social information is included in the recom-
mended locations and activities. According to [80], spatial
mendation system, the item can be labeled by the users.
data in large amount is unfeasible for usage and can be main-
Folksonomies are information spaces consisting of sets of
tained computationally for utilizing in planning procedures.
triples that specify a user, an item, and a tag [81–83].
Hence, the locations, distances, and driving directions are
obtained from geospatial web service technologies already
existing along with ESRI ArcWeb Service. Continuous cal- 4. Use of AI Techniques in Tourism
culation of user’s position and speed is made by GeOasis Recommender Systems
[53] to estimate the time required to reach a location so that
planning can be made in real time. Prediction is the key aspect Now, we brief the most prevalent AI methods that have been
of the system which determines whether the user would be exploited in recommender systems for tourism in the recent
on the road, in a city, or near a city. The planning algorithm years. Table 2 is a comparison of AI techniques used by travel
will consider only the nearest places if the users are in a city recommender systems in the articles reviewed for this paper.
already, assuming they are closer without counting the speed
or route. The attractions in a city are considered if the user 4.1. Multiagent Systems. Agents obtain information intel-
is near them. The algorithm gets complex due to temporal ligently from the environment in which they act upon
constraints when the user is on the move but far from a city. accomplishing the task or goals assigned. These software
The plan is computed by the client application not the server programs are proactive and automated. Optimal solutions
as constant checking of GPS to determine the location is to problems are obtained by multiagent systems where a set
needed. Google Maps acts as an external resource to compute of agents operate by coordinating and cooperating among
the routes. themselves in terms of resource and information sharing [88].
8 Computational Intelligence and Neuroscience

Table 2: Comparison of AI techniques used by travel recommender systems.

Management
Multiagent Optimization Automatic Knowledge
Articles of
system technique clustering representation
uncertainty
Garcı́a-Manotas et al.
✓
[84]
Sebastia et al. [63] ✓ ✓
Vansteenwegen et al.
✓
[44]
Fenza et al. [85] ✓
Garcia et al. [65] ✓
Gavalas and Kenteris
✓
[51]
Lorenzi et al. [59] ✓
Wang et al. [49] ✓ ✓
Batet et al. [67] ✓ ✓
Hsu et al. [86] ✓
Martı́nez-Santiago et al.
✓
[53]
Noguera et al. [58] ✓
Garcia et al. [69] ✓
Lucas et al. [72] ✓ ✓ ✓
Meehan et al. [73] ✓ ✓
Moreno et al. [87] ✓ ✓
Ruotsalo et al. [74] ✓ ✓ ✓

The Turist@ [67] implements a system based on agent service agent in order to query the databases for tourist infor-
in which suggestions for cultural activities are generated. An mation along with a personalization agent that implements
agent is assigned each of the concerned activities to maintain the CB methods to provide suggestion by selecting item based
a database of events related and its availability in the location. on the user profile and data.
Museum are each assigned an agent and hence are exempted. The PersonalTour [59] implements travel agents that are
A user agent provides an interface for interaction graphically. assigned to specific feature such as hotels, attractions, or
A broker agent interfaces the communication of information flights. Upon the arrival of a new user, the preferences and
in between the user agents and the agents assigned to each of interests are elicited which then are processed together with
the cultural activities. The user is provided with the ability the implemented agents to derive a suitable package for travel.
to interact in order to query, search, or evaluate recom- Evaluation of the package content and segments is possible by
the user, which in turn is regarded as an implicit feedback that
mendations. The recommender agent serves as the main
helps in improving the performance of the agent used and the
component of the Turist@ system by maintaining the tourist’s
suggestions provided by the system in general.
user profiles. Automatic-dynamic tuning of the knowledge
Castillo et al. [78], Lee et al. [43], and Sebastia et al. [63]
base is done which is obtained from the tourists by elicitation studies are some of the examples in which the system is built
during the initial usage. This initial knowledge comprises by the combination of agents that correspond to different
the interests and cultural activity preferences of the users. operational components such as the user interface, prefer-
Analysis and querying process are carried out in this initial ence and interest elicitation module, analysis and matching
knowledge in order to refine them. Proactive suggestions module, and route planning and generation module. In
are possible since the current user-location can be used to order to keep the system simple, communication and coordi-
determine the activities that match the user preference. The nation among the agents are kept minimal, which is achieved
system incorporates both CL and CB suggestion methods. by the sequential activation of the arranged agents. This
In Ceccaroni et al. [79], profile obtained initially from the limits the agents’ capabilities in terms of being concurrent,
user is modified by the analysis of both explicit evaluations distributed, and coordinated.
and implicit activity actions. A profile management agent
apart from initializing the profile by classifying the users 4.2. Optimization Techniques. Complicated plans and sched-
into stereotyped groups performs its modification based on ules have to be generated by the recommendation systems,
tourist feedbacks. This work proposes having an information which leads to a situation in which solutions to related
Computational Intelligence and Neuroscience 9

problems that are NP complete have to be optimally approx- 𝑘 Nearest Neighbors method is used to determine in
imated. Computational complexity is often considered as a which group a new user has to be added [90], which involves
trade-off in this regard; hence, suboptimal methods are also calculating the 𝑘 past users who are similar to the new one
considered in some cases. such as in [58, 91]. Once the new user has been allotted into
Ant colony optimization is used by Lee et al. [43], in the one of the groups, the suggestions can be given with respect to
design of a route recommendation system that incorporated the interests and preferences of the rest of the group members.
agents. Optimal coverage of the attractions along a route Domain ontology is analyzed to determine the similarity
is obtained here by approximating a solution of travelling index between the users of a group in SAMAP [78]. One
salesman problem by an autonomous entity set representing problem that has to be given attention here is scalability.
the ants that globally and indirectly communicate through Another method that is generally used to group similar
pheromone mediation. Genetic algorithm for the plan con- users is the 𝑘-means algorithm such as in Gavalas and
struction of a city tour is implemented in CT-Planner4 [71]. Kenteris [51] and Lucas et al.’s [72]. Here, 𝑘 is the number
It is an iterative approach that considers the possible plans as of clusters desired with respect to which the initial seeds are
the population with the user utility as the evaluation function. determined independent of the application. This is followed
The iterations are made after the mutation and crossover of by iteration of objects, sorting based on the calculation of the
the determined best population. The iterations lead to the nearest cluster, and recalculation of the cluster prototypes.
selection of the best plan evaluated with respect to the user Convergence of solution is achieved when repeated iterations
preference. In VISIT system [73] adaptive suggestions based place the objects under the same cluster. Application of 𝑘-
on the user context with respect to factors such as social means for determining an initial set of tour segments, classes
media sentiment, preference, weather, time, and location of similar users based on demography, and classification
are proposed. Artificial neural network is suggested for the with respect to explicit feedbacks is done in Moreno et al.
evaluation of the relevance between each context factor used [87]. 100 generic initial segments as types of users were
and the user profile. determined based on historical data that comprised 30,000
Vansteenwegen et al. [44] and Garcia et al. [69] are questionnaires. Each segment is related to a prototype and
some of the works that use heuristic based approaches level of preference with respect to each type of activity. Basic
to develop the travel routes. Another example is the City information is elicited from a new user by means of explicit
Planner system. Iterated local search can be also used in feedback forms which can be sufficiently used to place the
this kind of approaches by which generations of solution user in one of the groups and provide initial suggestions. The
sequences with respect to local search can be iterated. A composition of travel group, budget type of accommodation,
different solution to a route plan can be directed efficiently by and country of origin are the demographics considered in
the addition of heuristic after which the optimal solution can order to classify the user. In order to combine various kinds
be considered as the initial solution with respect to the local of data, operators of aggregation such as LSP [92] and OWA
search. Iteration can thus be continued till a threshold [93] are used.
criterion is achieved. In Souffriau et al.’s study [89] of Greedy In Fenza et al. [85], a variation of 𝑘-means in terms of
Randomized Adaptive Search techniques, query iterations uncertainty, the fuzzy 𝑐-means, is proposed, by which the
generate a set of possible locations or attractions that were object sets can be partitioned into clusters in such a way that
generated from the chosen start and end locations of the tour. every object membership degree lies between 0 and 1. And
The set thus obtained is filtered with respect to the heuristic for all clusters the addition of membership degree is 1. The
threshold value established initially and a random selection algorithm handles both the point of interest (poi) and the
is made on the remaining items. users. Once the POIs and the user clusters are defined rules
are derived, characterizing them so that the new user or POI
A∗ heuristic and hierarchical temporal planning were
can be placed in the best fitting cluster. The association rules
exploited in the SAMAP system [78]. Apart from these meth-
can also be built in such a way that they capture the relation
ods, recommender system for tourism is widely incorporated
between the POI and user clusters along with other pieces
with ad hoc planning methods in order to generate the routes
of information that are contextual. The rules obtained can
and plans that are personalized for a specific user. Also
then be used to calculate the activities and their types that
classical AI methods which are independent of domain are
can be suggested to a user. Similar methods can be found
also applied by some systems.
in the recommender system PSIS (Personalized Sightseeing
Information System) [72].
4.3. Automatic Clustering. CL methods in tour suggestion CL filtering methods that use class definition of grouped
systems are generally used to group the users based on the users are implemented in Turist@ [67]. Whenever the system
common attributes and features shared. Here, it is assumed gets an addition of 10 new users, clustering is redone to update
that it is appropriate to suggest to all other members of a the classes. In ClusDM [94], user interests and preference
group an activity or location that has been rated positively by along with the corresponding demographic data are used to
a user of the group. Information based on user demographics, build a class hierarchy. Desired number of classes can be
user preference, and rating history is used to determine these obtained by segmenting the tree thus obtained with respect
groups based on similarity by applying automated clustering to different levels.
techniques based on AI. We now further consider alternatives In SPETA system [95] classification by the usage of Sup-
of this approach. port Vector Machines (SVMs) is proposed. For this, storage
10 Computational Intelligence and Neuroscience

of user preference of various activities and its corresponding of such systems [97, 98]. Apart from these axioms and specific
characteristics in the form of vectors is made, upon which objects, other components are usually considered [99].
the SVMs can be used to determine the suitable and most In some of the recommender systems for tourism, ontol-
appropriate suggestions by analyzing the distance between ogy based formalization of the domain knowledge is made.
the item sets and the user preferences. Cultural activity suggestions are made by considering generic
ontologies in which information regarding various aspects is
4.4. Management of Uncertainty. Determination of suitable stored, such as Wang et al. [49], in which the formation with
suggestions to the users is difficult as the calculation of the respect to aspects such as restaurants, shopping, transport,
relation between the user demographics, preferences, and the accommodation, and culture comprises the generic travel
available POIs of a location involves complex methods. This ontology. The user ontology modeling considers the user
leads to uncertainty in the system which can be addressed by preferences and demographics. In GeOasis [53], SAMAP
AI method that involves approximate reasoning techniques [78], and SMARTMUSEUM [74], there is inclusion of ontolo-
that can determine and reason these uncertain relations. gies for modeling various kinds of user activities and items
Bayesian networks are one such possibility, in which and for semantic reasoning. Ontology was again used for
acyclic graph with representation of causality relations or similarity measuring and deduction of similar items or
internodal influences can be made in terms of the edges. groups which along with CL filtering methods were used to
Probability analysis is used to determine the possible or the generate the suggestions.
most appropriate parent for a node in case of its absence. In e-Tourism [45, 63, 65], methods based on ontology are
A probability table is used for this purpose. The table of used extensively. Domain ontologies which have the details
conditional probability consists of 2𝑛 nodes for a node that of various activities and events of a city are used. Ontology
has 𝑛 parents. This table indicates the chance of occurrence is used in SigTur that comprises concepts greater than 200,
based on the parent nodes presence or absence. Hsu et al. [86] hieratically arranged in 5 levels.
present a simple method that involves the Bayesian networks In SigTur, the storage of user preference is made in each
to determine the probability of POI to be preferred by a of the ontology nodes along with the confidence level of
user by considering various attributes such as nationality, the considered preference. The initialization with respect to
age, income, occupation, and travel purpose. The proposed these ontology references is elicited from the new user with
network considers that the probability that a user is likely to the help of a small questionnaire. The system is updated
prefer an activity or location is influenced by factors such as by the transmission and spreading of information from the
age, personality, and occupation. Activities or events are not children nodes to the parents when the user interacts with
specified in the used network. the system. The interaction may be a search, an addition
Fuzzy logic is also widely used to handle the uncertainty of certain activity to a generated plan, or any usage that
in various systems. Linguistic variable values are processed by provides additional knowledge about the user preference.
this method by operation on a series of values. The variables Hence, the user preferences are managed dynamically based
belong to a fuzzy set in which the corresponding values are on ontology [68]. Similarly in the e-Tourism recommender
mapped to a fuzzy membership function that results in values system, preference updates are made after the analysis of
between 0 and 1. Thus, a generalization of standard logic explicit user ratings.
is achieved. The fuzzy logic can be used to represent the An ontology set instead of integrated single ontology
user preferences and demographic and the possible item and has been proposed in some of the systems. In PaTac [79], a
activities list and map the possibility of a user choosing an separated ontology is maintained for activities and locations
item [43, 95]. The measure of how similar a user group or a like entertainment, restaurants, and hotels. The work pro-
particular user is with respect to other users or groups can poses the linking of ontologies based on user models based
be determined using the fuzzy logic analysis [72]. Also the on user stereotypes and the ontologies of W3C consortium
representation of contextual aspects pertaining to travel route which has temporal standard and is based on geolocation. In
or path can be done [73]. CONCERT [96], “ContOlogy” is used which is the inclusion
In some of the recommendation systems, fuzzy com- of 11 different ontologies that relate to attributes such as
ponent is avoided by the usage of rule based methods. services, tourism, preferences, motivation, or activities to
CONCERT system [96] involves suggestions generation by model an ontology network that considers all the travel
the rules that consider the user context and preferences. related contexts.
Usually the design and development of ontologies in the
4.5. Knowledge Representation. Representation of the domain recommender systems are manually built in an ad hoc man-
knowledge in tourism recommendation systems requires ner for a particular application. In [100], ontology population
methods that would be efficient and effective in inference is automated in order to minimize its construction cost. This
mechanisms such as in any knowledge or rule based system. is done by the analysis of electronic resources.
AI techniques are found to be adequate to represent and
build the knowledge base and to derive reasons from it.
In particular, ontologies are widely used nowadays for the 5. Location Based Social Networks
domain knowledge representation. Classes that represent the
concepts and described hierarchical relations that are taxo- More than one number of individuals connected together
nomical and nontaxonomical are the important components with more than one type of relations (e.g., friends, family,
Computational Intelligence and Neuroscience 11

common interests, and groups) is known as social network (ii) Point Location Based. Some applications like foursquare
[101]. A real world social network service can be digitally and Google Latitude mainly focus on people current loca-
represented. The social network not only mentions the users’ tions, such as hotel or park. Foursquare application is used
network, but also enhances their activities. The activities of a to point out the individual with the most number of check-
user depend on their actual ideas and on sharing posts and ins and a popular place with higher crowd ratings. The users’
events and making likes. real time location can be discovered by social network and
The user-location based social network data strengthens also this enables the social activities of the users in the real
the social network activities and also this location mentioned world. These real time location social activities are useful for
in the social network services. Location based social network inviting people to have dinner or go shopping with users.
formal definition is proposed by Zheng [102].
The location based social network comprises the people’s (iii) Trajectory-Based. The point locations and the route
physical location in their social structure to share the infor- connecting the point location are recorded by users and
mation by location embedded system. The new structure is are called trajectory-based social networking services (such
created when an individual user is connected to a location as Bikely, SportsDo, and Microsoft GeoLife). Normally, the
on a social network. The location of user derived from users’ experiences are represented by their tags, such as
their location tagged media content and other activities photos, media, and tips, along the trajectories and also
(such as their photos, video, and text). The user physical these services are used to record users basic information,
location consists of individual location at current time and such as distance, duration, and velocity. In addition to
their location history with specific period of time. If one or social networking services, trajectory-based service systems
more person has the same location and also similar location also provide the when and where information of users for
histories, it will not affect our social network structure. This personalization.
structure also contains individual behaviors, activities, and
other information.
The concept of locations based networks shows new 5.1. Unique Properties of Locations. Three unique properties
locations and correlations in addition to the old one. From of location based social network are hierarchical, measurable
the new information, graphs build into three types of location distances, and sequential ordering properties as shown in
based social network, such as location-location graphs, user- Figure 4.
location graph, and user-user graph.
5.1.1. Hierarchical Properties. Multiple scales in location span;
(i) Location-Location Graph. In this graph, users consecu-
that is, the location can be hotel or town and it depends on
tively visit the edge between two locations indicating the
users’ location. The different granularities of a location form
node location of the location-location graph. The correlation
a hierarchy; the locations with smaller area also connected
between strengths of two locations is represented by edge
with geographic areas. Each location has a relation with
weight.
another; for example, a hotel belongs to a neighborhood, the
neighborhood belongs to a town, and a town belongs to a
(ii) User-Location Graph. Users and locations are the two
country; location-location graphs and user-location graphs
types of entities in user-location graph. The visited location
have a location granularity in different levels (Figure 4(a)).
of the users is indicated by the edge starting from the users
The hierarchical relationships depend on users with lower
and ending at a location and the number of visits calculated
level location sharing and higher level location sharing.
by weight of the edge.
The lower level has a stronger connection than higher level
location sharing. This is a unique property of location based
(iii) User-User Graph. Basically a node is a user and edge social networks.
between two nodes represents two relations. The two relations
are existing social network between two users and a new
location of the users. 5.1.2. Measurable Distances. The new geospatial distance
relation has three types of location based social network
Three groups of location based social networking services that connected to physical world, different users locations
are geotagged-media-based services, point location based with their distance, user and location distance, and distance
services, and trajectory-based services. between two users locations. The first law of geography
stated that “everything is related to everything else, but near
(i) Geotagged-Media-Based. Locations are labeled to media things are more related than distant things” (Figure 4(b)).
content of users added by geotagging services. The new Location based social networks are affected by the influences
content of users is passively added to the physical world of similarity between user-user distances; for example, the
and also this content in the geographic context is viewed most visited location of user has more preferences [103], and
by the users. This location based social networking service the users are interested in particular location because the
is included in website (Flickr, Panoramio, and Geo-twitter). location is close to their homes, for example, hotel and park.
The social network services still focus on media content The correlation between locations also affects the location
because the connection between users is based on media based social networks because some places are close to each
itself. other.
12 Computational Intelligence and Neuroscience

State level Tamil Nadu D1

u2
City level Chennai ··· D2
Coimbatore u1

District level Kancheepuram Thiruvallur ··· D3 D4

(a) (b)

u1

u2

(c)

Figure 4: (a) Location hierarchy property. (b) Location distance property. (c) Location sequential property.

5.1.3. Sequential Ordering. The users’ chronological ordering the current location due to location sequential property. For
is used to mention the subsequent visits of two locations. example, more numbers of people visiting some important
For example, we consider the two users with their location place subsequently travel to the town.
visiting pattern shown in Figure 4(c). For each visit of users,
we can create an ordering of their similar location and 5.2.3. The Historical Locations of the User. The users’ prefer-
preferences. ence is indicated by the powerful histories of users’ behaviors
[104]. The LBSNs mention the user’s historical location and
5.2. Challenges to Recommendations in LBSNs. The new also reflect the user’s preferences, experiences, and living
LBSNs have three unique properties of locations. The unique patterns compared to the online behaviors of users. It is not
properties of LBSN are location context awareness, the heter- easy to model a location history of users because the loca-
ogeneous domain, and the rate of growth. tion history depends on distance, hierarchy, and sequential
properties of users. Based on the location history of users
we have to learn user’s personal preferences. Due to the
5.2.1. Location Context Awareness. What kind of recommen- following reason, it is very challenging work in LBSNs. (1)
dation system is needed for LBSNs to consider the users The challenging work is that we create users preference from
current location, users location history, and the influences of sparse location data because a full set location history of users
location histories to other users? does not exist. (2) The user’s location preferences are not only
limited to their hotel and shopping locations because user
5.2.2. The Current Location of a User. Due to the following has multiple kinds of interests: cycling, sports, movies, arts,
reasons, the current locations of users are more important and so forth. (3) Users preferences have granularity and also
parameter for generating recommendation system for LBSNs. follow some hierarchical steps like snakes → food → pizza.
The location granularity for different levels is represented by (4) The user’s preferences always depend on their location.
the current location of users. In recommendation system, it
is very difficult to choose a proper granularity. If we choose 5.2.4. The Location Histories of Other Users. Social opinion
hotel location of users that has a fine granularity, then a is one of the most important information bases for recom-
relative coarse granularity represents the town location of mended system making up with location history generated
users. by other users. From the location history we extract social
The most visited location is near to the users compared to opinions; it is not easy one because we are faced with the
the location at far distance; this implies the distance property following challenges. (1) The continuous representation of
of locations. But also the quality of location is important user’s changing location history is a complex task. (2) For
for making recommendation system for LBSNs because of each location, user has different knowledge. For example,
the ranking of recommendation system based on both the local user has expert knowledge to find high quality of hotel
quality of locations and the location close to users. Another and shopping malls. It is easy to interface user’s experiences
challenge is with respect to the collection of users’ fine and knowledge to the social opinion. From this users pref-
grain location, as it is frequently updated using mobile. By erence, we created a massive users location data. But for all
using efficient algorithms, the problem can be addressed with locations, the same users do not have this much knowledge
utilization of LBSNs. The future travel plan of users affects and location data.
Computational Intelligence and Neuroscience 13

Ratings
Recommendation prediction module
Data Methodologies User LBSN
objectives
sources interface data
module Recommendation
User module
Location recommender system

(1) User profiles (1) Content-based (1) Locations Figure 6: Proposed SPTW based location recommender system.
(2) User individual (2) Link analysis-based (2) Users
locations (3) Collaborative filtering (3) Activities
(3) User trajectories based (4) Social media

Figure 5: Categories of LBSN recommendation system, an over-

view. Set of users, set of location categories,
LBSN and ratings
User interface
module Prediction of trust pertinence
5.2.5. Heterogeneous Domain. The LBSNs heterogeneous
graphs consisting of user and location node and three edges User characteristics, location
are user-user, location-location, and user-location edges. The characteristics and weighted LBSN
LBSN is a model by at least three tightly associated graphs.
The trajectories of LBSN can be eliminated by another node of Perform social pertinent trust walker

Ratings prediction module

social network. A location is an important object in the LBSN Recommendation
and also it consists of users information. The connection module
Single solution per every iteration of
between two users is inferring heterogeneous graph of two trust walker
objects. The similarity between two users is obtained by their
location information sharing and will increase the strength of
Ratings calculation
connection between the two users.
LBSN Finalize the calculated ratings according
data
5.2.6. The Rate of Growth. The growth of location based social to the termination condition
networks is faster than normal social networks. LBSNs have
higher links, nodes, and social structure compared with the Figure 7: Overview of the proposed SPTW based location recom-
normal one. The social network for academic is consisting mender system.
of heterogeneous structure and conference with authors and
papers. For this social network, adding new links is more
difficult than LBSNs. Adding location to the LBSNs is easier
than publishing a paper in academic social network. The rate Social pertinent trust walker algorithm was introduced to
of growth of LBSNs rises with high efficiency and standard reveal the more relevant suggestions and make recommenda-
scalability and updating new location to the recommender tions more useful. The proposed recommendation model is
system. So, the research needs a year of experience regarding portrayed in Figure 6. The proposed location recommender
the users in LBSNs. system comprises four major components, namely, user inter-
Figure 5 shows the different types of social networks, such face module, ratings prediction module, recommendation
as social network for academic, social networks for general module, and location based social network.
online, and location based social networks. The proposed model can be restructured as in Figure 7.
In the proposed location recommender system, the user’s
interaction is done through the user interface module. Then,
6. Proposed Location Recommendation the user requests are forwarded to the ratings prediction
System Using Social Pertinent Trust module. Here, the ratings for the categories of locations were
Walker Algorithm predicted through proposed social pertinent trust walker
algorithm. The proposed algorithm is the extended version
In this section, we propose a social pertinent trust walker of random walk proposed by Jamali and Ester [105]. To
algorithm for an efficient location recommendation. In the predict the ratings for the location categories, location based
proposed recommendation model, locations that are rec- social network data is used. Then, the calculated ratings are
ommended to the user were predicted from the location forwarded to the recommendation module, where the ratings
based social network. Social pertinent trust walker algorithm are exploited to make list of places as recommendations along
determines the rating score of the locations based on the with the help of LBSN data.
existing score rated for the similar location categories. After The main ingredient of this work is trust between the
computing the rating score for the location categories, the list users of a location based social network. Figure 8 is rep-
of locations with more relevance is recommended to the user. resentation of location recommendations based on trust
14 Computational Intelligence and Neuroscience

The trust relevancy between two users 𝑥 and 𝑦 can be

defined as
u1 u2 u3 TruRel = SimUser (𝑥, 𝑦) ∗ Trust (𝑥, 𝑦) , (1)
LBSN
where SimUser(𝑥, 𝑦) is the similarity between users 𝑥 and
𝑦 and Trust(𝑥, 𝑦) is the degree of trust of user 𝑥 to user 𝑦.
u4 u6 The degree of trust between the users is calculated using the
u5
existing algorithms based on the historical interactions of the
users [106]. Through calculating the trust relevancy between
the connected users of location based social network, the
weighted LBSN can be obtained. In order to calculate the
Culture Family Friends Lifestyle Spiritual similarities between the users, matrix factorization is used. To
c1 c2 c3 c4 c5
define matrix factorization, the user-location category rating
Locations categories matrix 𝑅 can be decomposed into matrices, namely, 𝐸 and 𝐹,
Figure 8: Location recommendations based on trust enhancement
in LBSN. 𝑅 ≈ 𝐸𝐹𝑇 . (2)

The similarity between two users 𝑥 and 𝑦 can be calcu-

lated as follows:
enhancement in LBSN. Every user is capable of invoking 𝑥⋅𝑦
with different categories of locations and rates the location SimUser (𝑥, 𝑦) = cos (𝑥, 𝑦) = 󵄨 󵄨. (3)
categories based on their experiences. In a situation, while |𝑥| ⋅ 󵄨󵄨󵄨𝑦󵄨󵄨󵄨
a user is requesting a location recommendation, the ratings
for the particular user are calculated and the locations with 6.2. Social Pertinent Trust Walker Algorithm. In the proposed
the higher predicted ratings are recommended to the user. work, the SPTW algorithm has been designed to discover
Therefore, utilization of LBSN data with ratings enhances the the interesting category of locations for the particular user
from the location based social network. The SPTW algorithm
accuracy of recommendations.
reaches final solution after several iterations. For every
Generally, the user set 𝑈 = {𝑢1 , 𝑢2 , 𝑢3 , . . . , 𝑢𝑘 } and iteration, the random walk initiates from the target user
location categories set 𝐶 = {𝑐1 , 𝑐2 , 𝑐3 , . . . , 𝑐𝑛 } are the major 𝑈𝑥 in a weighted LBSN. While reaching the 𝑛th step of
components of a trust based location recommender system. random walk in a trusted LBSN, the 𝑢𝑠𝑒𝑟 is reached by
The user rating is expressed through a matrix called rating the process. If 𝑢𝑠𝑒𝑟 has rated the location category to be
matrix and it can be denoted by 𝑅 = [𝑅𝑢,𝑐 ]𝑘×𝑛 . The ratings recommended, 𝐶𝑆𝑥, then this rating is used as 𝑟𝑎𝑡𝑒𝑥 for
can be any integers within a range. Mostly, the range of the that iteration. Else the random walk will be terminated based
ratings is between 1 and 5. One denotes poorly liked and on the 𝑅𝑎𝑛𝑑𝑜𝑚 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝑇𝑟𝑢𝑠𝑡𝑊𝑎𝑙𝑘𝑢𝑠𝑒𝑟,𝐶𝑆,𝑛 . On the
five denotes heavily liked. In a location based social network, other hand, the trust walker continues its walk based on
the trust between two users is denoted by real numbers 0, 1. 1-𝑅𝑎𝑛𝑑𝑜𝑚 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝑇𝑟𝑢𝑠𝑡𝑊𝑎𝑙𝑘𝑢𝑠𝑒𝑟,𝐶𝑆,𝑛 , one when the
The representation of trust is that one means full trust and target user trusts the next node of weighted LBSN.
zero means no trust. The trust between users is expressed Algorithm 1 is a trust walk model for calculating rating
through a trust matrix. In the trust matrix 𝑇 = [𝑇𝑢,V ]𝑘×𝑘 , for category of location for the specific user and it is eval-
the relationship between the users or the trust between the uated using real time location based social network dataset
users is denoted by nonzero value. Finally, the problem of (foursquare). The results prove that the algorithm is optimal
location category recommendation can be defined as follows: and shows better performance through the evaluation met-
user 𝑢 requests a category of location 𝑐 for which ratings rics such as coverage, precision, and 𝐹-measure. Compared to
are unidentified, hence the prediction of ratings 𝑟 for user 𝑢 the existing trust walk models, the proposed algorithm shows
though utilizing the ratings matrix 𝑅 and trust matrix 𝑇. betterment in accuracy of recommendations. Based on the
recommended category of locations from the proposed social
pertinent trust walker algorithm, the list of top 𝑘 nearest
6.1. Trust Relevancy. It is very well known fact that the trust locations will be suggested to the user.
relationship between the users of LBSN will not directly help The working of the SPTW algorithm is demonstrated in
in the enhancement of accuracy of recommendations. But the Figures 9(a) and 9(b). The edge between the users holds the
categories of locations recommended by the trusted users can probability as a weight. Let us consider location category
be considered to be more reliable. Through this contradiction, 𝑐4 which is considered to be recommended to the user 𝑢1 . In
the recommendation of location categories may be affected the first step of trust walk as shown in Figure 9(a), the next
in the prediction of ratings for the particular user due to node/target node is chosen as 𝑢4 based on larger probability
difference in preferences, interests, and observation. Hence, edge value. If user 𝑢4 has rated the location category 𝑐4 with
considering both similarities between the users and trust the rating ratex, the result of the trust walk will return as
relationship may be more helpful to address the above ratex. If the termination condition has not been reached,
challenge. the iteration continues for the second walk. In this step,
Computational Intelligence and Neuroscience 15

Input: 𝑈 (set of users), 𝑈𝐿 (User Location), 𝐶𝑆 (set of location categories), 𝑅𝑀 (Rating

Matrix), 𝐿𝑆𝐵𝑁 (weighted location based social network), 𝑈𝑥 (specific user/target user) 𝐶𝑆𝑥
(to be recommended category of location)
Output: 𝑅𝑎𝑡𝑒𝑥 (predicted rating for category of location)
assign 𝑛 = 1; // step of the trust walk
assign 𝑢𝑠𝑒𝑟 = 𝑈𝑥; // assigning the initial point of the trust walk as 𝑈𝑥
assign extreme depth = 6; // allocation of maximum steps for trust walk
assign ratex = 0;
if (𝑈𝐿 ≠ 0)
{
CUL = user location category (𝑈𝐿, 𝐶𝑆)
}
while (𝑛 <= extreme depth)
{
user = ChooseUser(user); //choose next node from Trusted User Set 𝑢𝑠𝑒𝑟
as next step’s target as per the probability calculated by the
function 𝐸user (next node)
if (𝑢𝑠𝑒𝑟 has rated 𝐶𝑆//𝐶𝑆 == 𝐶𝑈𝐿)
{
ratex = ratex𝑢𝑠𝑒𝑟,𝐶𝑆 ;
return ratex;
}
else
{
if (random(0,1) < Random Probability of TrustWalk𝑢𝑠𝑒𝑟,𝐶𝑆,𝑛 //
𝑛 == extreme depth) //end at present node
{
𝐶𝑆𝑥𝑖 = choose location category(user); //location category CS𝑥𝑖 is choosed from
Rated Location Category𝑢𝑠𝑒𝑟 based on the
probability of Rating probability𝑢𝑠𝑒𝑟 (CS𝑥𝑖 )
ratex = ratex𝑢𝑠𝑒𝑟,𝐶𝑆𝑥𝑖 ;
return ratex;
}
else
𝑛++;
}
}
return ratex;

Algorithm 1

0.5 1 0.5 1
u1 u2 u3 u1 u2 u3

1 1 1
0.7 0.4 0.7 0.6 1

0.8 1 0.6 1
u6 u6
u4 u4
u5 u5
(a) (b)

Figure 9: (a) Working of social pertinent trust walker (Step 1). (b) Working of social pertinent trust walker (Step 2).

user 𝑢5 is chosen as specified in Figure 9(b). Again the from the list of location categories the user 𝑢5 has rated.
algorithm checks whether the user 𝑢5 rated the location The probability of selecting the similar service is based
category 𝑐4 . If the user 𝑢5 has not rated the location category on 𝑅𝑎𝑡𝑖𝑛𝑔 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑢𝑠𝑒𝑟 (𝐶𝑆𝑥𝑖 ). The rating of the chosen
𝑐4 , but the termination condition has been achieved, then similar location category rated by the user 𝑢5 will be assigned
the algorithm chooses the most similar location category as ratex and will be returned as a result for the iteration.
16 Computational Intelligence and Neuroscience

7. Group Recommendation
The proposed SPTW recommendation model is extended to
group recommendation for the members of the particular Group
group. The popularity of the POI plays a main role in the
group recommendations. The proposed group recommenda- Group profile generation
tion model suggests POI to the group members who appeal
for the travel recommendation. The proposed model is on
Recognize location/POI
the basis of the location category of the particular POI and category for group
the popularity of POI is used to find relevant location for the
group.
In general, group members may be homogeneous or het- Popularity of POI Consideration score of POI
erogeneous based on user preference regarding the locations.
SPTW based group recommendation model (SPTW-GRM)
exploits the interests of every user of the group along with Compute general ranking of
the interested location category from their social profile. each POI based on
Interested location category with its repetition in the user location category
profiles of the group members is used to form a group profile.
The created group profile replicates the common interested
location category with higher rating of the user’s group.
Later, based on location category relationship attributes, the
proposed recommendation model determines POIs of the
datasets whose exact category is not yet assigned. Then, the
general ranking of the POIs is computed through proposed
collective ranking function for the particular group. The
computed general ranking of POI is based on the popularity
Top-n POI recommendations for the group
of POI and the consideration score of POI for the group.
The popularity of the POI represents the opinion of all Figure 10: Proposed SPTW based group recommendation model.
users and their feedback regarding the particular POI. The
consideration score of the POI reflects the repetition of the
location category presented in the group profile. The process 7.2. Location Category Relationship Attributes. SPTW-GRM
and working of the SPTW based group recommendation depends on the location category relationship attributes to
model are illustrated in Figure 10. calculate the similarity between two location categories.
The similarity computation process helps to identify more
7.1. Group Profile Generation. The main aim of the proposed relevant POIs that should be considered to be part of top-𝑛
SPTW based group recommender system is to generate list of list. The location category relationship attributes are also used
POIs to the group of users through analyzing the preference to determine the consideration score for the POI concerning
of every member of the group. The system creates a group particular group. The location category relationship attributes
profile that combines all users’ preferences together to form are calculated extensively by the similarity calculation process
group preferences. The group profile creation is adapted from of the proposed SPTW algorithm. SPTW-GRM has adopted
an aggregation model of merging individual users into a the similarity calculation model used by the SPTW algorithm
group [107, 108]. The group reflects the combinational inter- and it classifies the location/POI based on its features.
ests of group members with respect to location categories.
The proposed model finds the interesting location cate- 7.3. Discovering Relevant POIs to Be Recommended. The
gory for the particular user of the group to assign relevant number of POIs available on LBSN is high and to find a
location category to the user. Personal location category is relevant POI in the vast list is a complex task. Analyzing
classified by the user and it represents the user’s interests each POI with respect to interests of group members is not
towards POIs. Then, SPTW-GRM creates a common group an efficient way to generate recommendations. To reduce
profile to the members of the group and the profile includes the number of comparisons, SPTW-GRM follows a specific
the common location category assigned by the group mem- filtering mechanism to create a set of location categories
bers in their individual profiles. The location category with that has to be considered for POI recommendation. The
the higher repetition in the group profile shows the users’ specific filtering mechanism of the SPTW-GRM considers
interests towards that location category and it has more the personal location categories of the group members and
impact on POI recommendations compared to the location includes the most similar location category with the list. Since
categories with lower repetition. Figure 11 shows the creation the location category describes the POI, the location category
of group profile by SPTW-GRM considering the location can reflect the interest of the group too. The recommended
category available on the individual user profiles of group POIs are the outcome of group preferences on location
members. category used by the POI.
Computational Intelligence and Neuroscience 17

POI 1: culture, family Culture, family,

POI 2: friends, entertainment friends, entertainment,
··· ···
Member 1
Culture, family(2),
POI 3: lifestyle, friends Lifestyle, friends(3), friends(5), entertainment(2),
POI 4: leisure, friends leisure, adventure, lifestyle, leisure,
POI 5: adventure, friends adventure, spiritual,
Member 2 ··· ···
···
POI 6: spiritual, friends Spiritual, friends,
POI 7: entertainment, family entertainment, family,
··· ···
Member 3

Location category assignments Individual user profiles Group profile

Figure 11: Group profile creation for the members of the group based on the location category of POI.

7.4. Group Recommendation Generation. After discovering group members are interested. The CScore of POI for group
relevant set of POIs based on location category, which is to be 𝐺 is described as
recommended to the group of users, SPTW-GRM proceeds to
compute ranking for the POI based on popularity of the POI CScore (POI, 𝐺) = ∑ ∑ LCRA (lc, poi)
lc∈𝐺 poi∈POIL
and consideration score for POI. Both features are combined
together to form top-𝑛 POIs as recommendations. It is to Repetitionlc
be noted that top-𝑛 recommended POIs will have higher × (4)
Max (Repetition𝐺)
consideration scores with respect to the particular group
taken into consideration. Repetitionpoi
× ,
Max (RepetitionPOIL )
7.4.1. Popularity of POI. The popularity of the POI is used
to find the rating of it by global users. Through this con- where LCRA(lc, poi) is location category relationship
sideration, the users’ satisfaction levels are obtained through attributes of location category lc and point of interest (poi).
the WoM (Word of Mouth) and feedback systems. The The Repetitionlc is the repetition of location category in the
communal interests on the POI are generally expressed by group profile 𝐺 and Max(Repetition𝐺) denotes the maximum
the users as ratings on LBSNs. The highly rated POIs on repetition in the group profile 𝐺. The Repetitionpoi is the
the LBSN have higher ranks when it is considered for the representation of repetition value of particular POI and
recommendation. Popular POIs attract more users and have Max(RepetitionPOIL ) denotes the maximum repetition in
more check-ins on LBSNs. The SPTW-GRM analyzes the the POI list. The weight of Repetitionlc /Max(Repetition𝐺),
check-ins and ratings of the POIs based on group users’ Repetitionpoi /Max(RepetitionPOIL ) is used for location
interests and location category considered. If the popularity category and point of interest, respectively. Repetitionlc is
of the POI is alone considered for the recommendations, then the indicator that represents the user’s interests on location
the relevance degree of the recommendations may be low category in the group. Larger value of Repetitionlc shows
and earlier research shows that such recommendations are more importance of location category to the group. SPTW-
useless. To enhance the accuracy of the POI recommenda- GRM considers all these weighted attributes as consideration
tions, the popularity of the POIs should be considered along score while generating top-𝑛 POI as recommendation.
with the user’s personalized needs. The popularity of the POI
which is considered by SPTW-GRM is a decision factor to 7.4.3. General Rank Computation. After calculating the over-
generate recommendations for all users who have check-ins all popularity and consideration score of each POI, SPTW-
at POIs of LBSNs. GRM estimates the general ranking for the POI through the
CombMNZ, a linear combination measure proposed by Lee
[109]. CombMNZ is used in many data fusion models and it
7.4.2. Consideration Score of POI. In parallel to the POI popu-
takes multiple rakings of item 𝐼 into consideration to calculate
larity computation, SPTW-GRM calculated the consideration
combined raking of item 𝐼. The CombMNZ model for POI
score for POI with respect to the group. The degree of users’
can be represented as follows:
interest in a group for the POI is determined by SPTW-
GRM by calculating the consideration score. The consid- 𝑛
󵄨 󵄨
eration score CScore(POI, 𝐺) collects the location category CombMNZPOI = ∑ POIrl × 󵄨󵄨󵄨󵄨POIrl > 0󵄨󵄨󵄨󵄨 , (5)
relationship attributes for all location categories in which rl=1
18 Computational Intelligence and Neuroscience

where the number of ranked lists to be merged is represented are termed as venues and the visits are called check-ins.
by 𝑛. In other words, 𝑛 is the number of ranked lists taken Foursquare is a real dataset that consists of user ratings for
as inputs. The computed normalized score for all nonzero venues and check-ins which is used to evaluate the quality
POIrl of the POI for the ranked list rl is to be merged. Before of the recommendations made (Table 3). The foursquare data
calculating the general ranking score of the POI, the ratings is used with our experimental setup with the extraction
of the individual rl should be converted into a range using the of user’s social connections. This dataset contains 2153471
following equation: users, 1021970 check-ins, 1143092 venues, 2809581 ratings,
and 27098490 trust relations which has been extracted from
RatingsPOI − POIrlmin the foursquare application through the public API. As there
POIrl = , (6) is no classification of users based on gender in the dataset,
POIrlmax − POIrlmin
we have adopted a classification model to classify the users
anonymously into male and female for evaluation purpose.
where RatingsPOI is the ratings of the POI in the rl and POIrlmax
The possible user data from the dataset is classified into male
and POIrlmin represent the maximum and minimum rating of users and female user sets.
POI in the rl.
The evaluation of social pertinent trust walker algorithm
The SPTE-GRM exploits the popularity of the POI and
is done through comparing it with the following state-of-
consideration score of the POI calculated in the earlier
the-art methods. Golbeck [110] proposed a TidalTrust model
sections and it normalizes the value through the above
which generates the ratings for the user through a trust
equation. Then, the value of 𝑛 is assigned to be 2 in
inference algorithm. Massa and Avesani [111] have designed
CombMNZ calculation, since there are only two input sets,
a model called MoleTrust, by which the trust score for the
that is, popularity score and consideration score. The general
target user will be predicted through walking along social
ranking score of the POI is determined and the SPTW-GRM
network. TrustWalker is a random walk model proposed
recommends top-𝑛 relevant POI to group members. Through
by Jamali and Ester [105] which utilizes the similarity and
utilization of data merger model, the proposed work explores
trust. RelevantTrustWalker is an extension of TrustWalker
the efficiency of consideration scores and popularity scores of
introduced by Deng et al. [112] that uses the degree of trust
the POI for efficient recommendations with higher popularity
between the users to predict the ratings.
and relevance.
The evaluation of SPTW-GRM for POI recommendation
for group members is done on foursquare dataset. As the
8. Experimental Evaluations dataset is on individual users, we have created a modified
In this section, we show the results from the experiments dataset for the evaluation process. As there is no benchmark
to evaluate the performance of social pertinent trust walk dataset available for POI recommendation for group users,
algorithm through foursquare dataset. The proposed model we have adapted familiar approach to make group of users for
is implemented in Java JDK 1.7 on Intel Core i7 3.1 GHz evaluation process. The generation of user groups is done on
machine with 16 GB of memory running Microsoft Windows basis of uniformity and size of the group. We have adapted
7. Upcoming subsections describe the dataset and evaluation group sizes from 2 to 8 to reach higher consensus between
methodology with discussion of experimental results. group members [113]. Based on similarity between users,
the groups are classified into three types, namely, random,
8.1. Dataset. Foursquare is a location based social network dissimilar, and highly similar groups. The similarity between
which holds user’s previous visits to locations. Locations users is calculated as

󵄨󵄨 󵄨 󵄨 󵄨
󵄨{poi | poi ∈ POI𝑢 ∧ poi ∈ POIV ∧ 󵄨󵄨󵄨rating (𝑢, poi) − rating (V, poi)󵄨󵄨󵄨 ≥ 2}󵄨󵄨󵄨
Similar Users (𝑢, V) = 󵄨 󵄨󵄨 󵄨 , (7)
󵄨󵄨{poi | poi ∈ POI𝑢 ∨ poi ∈ POIV }󵄨󵄨󵄨

where POI𝑢 represents the set of POIs rated by user 𝑢 and 8.2. Evaluation Metrics. We adopt benchmark evaluation
poi is a point of interest and POIV represents the set of POIs metrics to assess the performance and accuracy of the pro-
rated by user V. |rating(𝑢, poi) − rating(V, poi)| ≥ 2 is the posed SPTW recommendation method and the results were
limitation of a poi shared by both users 𝑢 and V. As the ratings compared with the baseline recommendation algorithms.
are of traditional rating scale from 0 to 5, the similarity may Generally, the accuracy of the recommendation algorithms is
be considered as high and low. Based on the similarity score analyzed by root mean square error and we have also adopted
computed by Similar users(𝑢, V), the users of the foursquare RMSE to determine the error in generated recommendations
dataset have been grouped as highly similar and dissimilar
2
groups with group size ranging from 2 to 8. We have also ̂
√ ∑𝑢,𝑐𝑠 (𝑅𝑢,𝑐𝑠 − 𝑅𝑢,𝑐𝑠 ) (8)
created random groups of users for the evaluation purpose. RMSE = ,
The random group may contain similar users or dissimilar 𝑁
users, but its size of users also ranges from 2 to 8. The groups where 𝑅𝑢,𝑐𝑠 is the real rating user 𝑢 has given to the category
are formed to hold uniform cohesiveness. of location 𝑐𝑠 and 𝑅 ̂ 𝑢,𝑐𝑠 is the predicted rating of user 𝑢
Computational Intelligence and Neuroscience 19

Table 3: Foursquare dataset statistics. Here, DCG∗ is the ideal DCG, where pois are arranged
in the decreasing order till the position of 𝑘 with respect to
Features Gross units
𝑅𝑢,poi .
Users 2153471 For a list of 𝑘 points of interest, an average NDCG is
Venues 1143092 defined as follows:
Check-ins 1021970 𝑢
Ratings 2809581 1 0
NDCG@𝑘 = ∑ NDCG@𝑘 (𝑢) . (14)
Trust relations 27098490 𝑢0 𝑢=1

MAE is the Mean Absolute Error used to evaluate the

to the category of location 𝑐𝑠. The total number of tested proposed model and it is defined as
ratings is denoted by 𝑁. The larger value of RMSE denotes 󵄨 󵄨
the inaccuracy of the recommendation algorithm whereas ∑(𝑢,𝑝𝑜𝑖)∈𝑅test 󵄨󵄨󵄨󵄨PR𝑢,poi − AR𝑢,poi 󵄨󵄨󵄨󵄨
MAE = 󵄨󵄨 󵄨 , (15)
the smaller value of RMSE represents the preciseness of the 󵄨󵄨𝑅test 󵄨󵄨󵄨
algorithms.
Due to the high sparsity of the dataset used, some where PR𝑢,poi denotes the predicted ratings for user 𝑢 for poi
recommendation algorithms are not capable of predicting and AR𝑢,poi represents the actual ratings and 𝑅test is the set of
the ratings. Utilizing the trust between the users increases users and point of interests used for test purpose.
the coverage and precision ratio. Coverage metric is used
to calculate the predicted value of generated pair of the 8.3. Experimental Results. The results of the various experi-
user and category of locations. In any circumstances, if the mental comparisons with recommendation results and per-
recommendation algorithm is not capable of predicting the formance evaluation of different algorithms are presented in
similarity between the user and category of locations, then it this subsection. From the experimental results of Table 4 and
is said that the algorithm did not cover the particular pair of infographics of Figures 12(a), 12(b), 12(c), and 12(d), it is clear
user and category of location: that there is no complete gain for trust based algorithms in
terms of recommendations accuracy. Though other models
𝐶𝑆
Coverage = , (9) consider trust of the users for recommendations, the pro-
𝑁 posed SPTW combines both trust of users and similarity
where 𝐶𝑆 denotes the number of ratings predicted and 𝑁 between users along with the current location. By this, the
denoted the number of ratings tested. accuracy of SPTW has been improved a lot compared to
RMSE and coverage can be combined into single metric other methods and the performance of recommendation is
as 𝐹-measure. To calculate 𝐹-measure, it is necessary to better compared to other algorithms. Obviously, SPTW has
determine precision values. Precision values can be obtained less RMSE compared to other algorithms due to the selection
from RMSE. For such calculation, RMSE value should be of appropriate users based on trust and similarity and the
converted into precision metric range of 0 to 1: coverage of SPTW is also high among other algorithms.
While considering the metric of 𝐹-measure, SPTW performs
RMSE very well along with other algorithms considered for com-
Precision = 1 − . (10)
Maximum Possible Error parison. Due to enormous size of dataset, the time taken
for location recommendation should be also considered as
In the above equation, RMSE is divided by the maximum an evaluation factor. Figure 12(e) portrays the average time
possible error, that is, 4. Here, 4 is considered as maximum cost consumed by various algorithms taken for consideration.
possible error due to the ratings in the range of 1 to 5. Since trust calculation is a bit additional task added to the
𝐹-measure is determined by using coverage and precision prediction and recommendation process, it makes an impact
as follows: over computational time cost. Compared to other models
2 × precision × coverage that consider trust between the users, SPTW costs less time
𝐹-Measure = . (11) because, in every iteration or walk, SPTW chooses target
precision + coverage
node based on probability and trust relevancy. Such an
NDCG is the normalized version of DCG (Discounted approach makes SPTW algorithm more reliable and helps to
Cumulative Gain) which is defined as make recommendations more quickly through the improved
computational efficiency.
𝑘 𝑔𝑢,poi(𝑗) The evaluation of the SPTW algorithm and its POI
DCG@𝑘 (𝑢) = ∑ , (12)
𝑗=1 max (1, log𝑏 𝑗)
recommendation score were done as assessing a traditional
information retrieval system. Hence, we have adapted famous
where 𝑔𝑢,𝑖 is the gain of each user (𝑢) when point of interest evaluation metrics NDCG (Normalized Discounted Cumu-
(poi) is recommended and 𝑗 is the position of the poi in the lative Gain) and MAE (Mean Absolute Error). NDCG is
recommendation: used to determine the performance of the proposed model
through analyzing the relevant POIs in the recommendation
DCG@𝑘 (𝑢) list. MAE is used to find the number of errors in the generated
NDCG@𝑘 (𝑢) = . (13)
DCG∗ @𝑘 (𝑢) recommendations. The large value of MAE shows the higher
20 Computational Intelligence and Neuroscience

Table 4: Comparison and performance evaluation.

Algorithms RMSE Coverage (%) Precision 𝐹-measure Time cost

TidalTrust 1.247 83.25 0.6883 0.6826 210.34
MoleTrust 1.193 87.58 0.7018 0.6967 255.48
TrustWalker 1.122 90.23 0.7195 0.7138 549.45
RelevantTrustWalker 1.086 92.47 0.7285 0.7228 510.49
SocialPertinent 1.041 95.78 0.7398 0.7341 459.29
TrustWalker

error with the recommendations. Figures 13(a), 13(b), and the average processing time of group users. We have also
13(c) portray the NDCG comparison of all, male, and female evaluated all types of user sets to determine the efficiency
users, respectively. The tuning parameter 𝑘 plays a key role through average processing time. The comparison of average
in the performance of SPTW algorithm. The weights of processing time of SPTW-GRM for different types of user
the tuning parameter 𝑘 are used to identify the optimal sets is portrayed in Figure 15(b). The average processing
performance of SPTW algorithm in various datasets. The time shows the time taken for recommendation generation
optimal performance is obtained in all users and male dataset depending on group size and uniform levels between the
when 𝑘 = 0.5 for all NDCG@5, 10, and 20. In the female users users. The results of the SPTW-GRM show that one user can
dataset, the optimal performance is obtained when 𝑘 = 0.6 for influence another user’s opinion. In other words, satisfaction
NDCG@5 and 𝑘 = 0.5 for NDCG@10, 20. The comparison levels of one user can have an impact on another user of the
of MAE for male, female, and all users datasets is depicted group through the generated recommendation.
in Figure 13(d). As the behavior of the users’ choice selection
randomly differs, it is also to be considered that the dataset
is very sparse. The sparsity of the dataset can be addressed 9. Conclusion
using the tuning parameter. The results of the MAE show that
when the tuning parameter 𝑘 = 0.5, SPTW shows optimistic This final section is the summary of the work presented in this
results and it can be followed in the future experiments and paper, which describes the key points that should be taken
evaluations. into consideration by the researcher, who is aiming to develop
The group recommendation based on SPTW-GRM is a recommender system. This paper creates an impact through
evaluated on the uniformly distributed group with the sizes the outline of several future work challenges in the area of
ranging from 2 to 8. The performance of the SPTW-GRM recommender systems’ design and development. Through the
is evaluated using NDCG and MAE for performance and analysis of interfaces used by the recommender systems, it
efficiency. SPTW-GRM is implemented on the formed groups is very well noticed that the recent development of mobile
on foursquare dataset based on similarity between users as platforms has been utilized very little. Clever exploitation
highly similar, random, and dissimilar groups. Popularity of of mobile platform with the personal data such as current
POI and consideration score increase the accuracy of the location may help in providing precise recommendations to
groups. The complete comparison of NDCG is represented in users in an improved manner. Most of travel recommender
Figures 14(a), 14(b), and 14(c) for highly similar, random, and systems lack the points of personalization, interactivity, and
dissimilar groups, respectively. The comparisons of MAE for adaptivity. Though TRS provide points of interests as their
highly similar, dissimilar, and random users are portrayed in suggestion according to the user preferences, the system still
Figure 14(d). The difference between the results of the SPTW- needs user’s help to build their trip manually. Some research
GRM for various groups with various users’ sizes shows the has tried to solve the issue of automaticity in travel planning
statistical significance of the work. From the results, the service, but still the problem of automatic travel planning
observations are done to determine the uniformity of the is yet to be addressed. This is a novel issue where social
group members in the various groups. The highly similar information and context of the user can be utilized to solve
groups hold higher uniformity levels compared to other the problem.
groups. The recommendations for this group are generated Every classical approach (such as collaborative, content-
quickly compared to the other two types. The random group based, and demographic) suffers from various problems in
comes next in the performance and the dissimilar groups providing personalized recommendations to the individual
hold last position in the performance, efficiency, and accuracy user. The trend of hybrid recommendation models along with
of recommendations. As a next step in the evaluation of the contextual information of the user may solve such issues
the SPTW-GRM, the efficiency of the model is analyzed for of individuality. Usage of the implicit and explicit preferences
various group sizes regardless of its type. The processing time of users extended with the semantic models addresses the
is computed for all group sizes and average is taken. The problem of uncertainty in the recommendation process. An
average processing time increases as the size of the group artificial intelligence technique such as knowledge represen-
increases. The trend of average processing time denotes the tation is commonly used for reasoning the recommendation
scalability of the SPTW-GRM and Figure 15(a) represents process. Automatic clustering algorithms may be used to
Computational Intelligence and Neuroscience 21

1.5 100

1.2 80

0.9 60

Coverage
RMSE

0.6 40

0.3 20

0.0 0
TidalTrust MoleTrust TrustWalker Relevant- SocialPertinent TidalTrust MoleTrust TrustWalker Relevant- SocialPertinent
TrustWalker TrustWalker TrustWalker TrustWalker
Recommendation techniques Recommendation techniques
(a) (b)
1.0 1.0

0.8 0.8

F-measure
0.6 0.6
Precision

0.4 0.4

0.2 0.2

0.0 0.0
TidalTrust MoleTrust TrustWalker Relevant- SocialPertinent TidalTrust MoleTrust TrustWalker Relevant- SocialPertinent
TrustWalker TrustWalker TrustWalker TrustWalker

Recommendation techniques Recommendation techniques

(c) (d)
700

600
549.45
510.49
500 459.29
Time cost

400

300
255.48
210.34
200

100

0
TidalTrust MoleTrust TrustWalker Relevant- SocialPertinent
TrustWalker TrustWalker
Recommendation techniques
(e)

Figure 12: (a) Comparison of RMSE. (b) Comparison of coverage. (c) Comparison of precision. (d) Comparison of 𝐹-measure. (e)
Comparison of time cost utilized by different algorithms.

classify users and optimization techniques can be deployed information from the user is the exploitation of their social
to generate the cost effective recommendations to the user. network data. Since tourism domain is very social in nature,
The problem of complexity in the scheduling and planning this data may help much in the development of recommender
process of route generation is a novel problem in the travel systems based on user interests. Utilization of the social data
route recommendation. The best way of obtaining more such as check-in behavior, ratings, social relationships, and
22 Computational Intelligence and Neuroscience

1.0 1.0

0.8 0.8

0.6 0.6

NDCG
NDCG

0.4 0.4

0.2 0.2

0.0 0.0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
k-parameter k-parameter
NDCG@5 NDCG@5
NDCG@10 NDCG@10
NDCG@20 NDCG@20
(a) (b)
1.0 1.5

0.8 1.2

0.6 0.9
NDCG

MAE

0.4 0.6

0.2 0.3

0.0 0.0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
k-parameter k-parameter
NDCG@5 All users
NDCG@10 Male users
NDCG@20 Female users
(c) (d)

Figure 13: (a) Comparison of NDCG for all users. (b) Comparison of NDCG for male users. (c) Comparison of NDCG for female users. (d)
Comparison of MAE for male, female, and all users.

recent area of work can help in the discovery of more accurate recommender systems were focusing on filtering mecha-
travel recommendations, which fits better the tastes of the nisms to improve the accuracy of recommendations. Now,
user. hybrid algorithms incorporated with the various factors-
For every recommender system, it is very important to influenced data have been taken into consideration in the
hold specific information about users and their interests as development of efficient recommendation models.
a profile. The development of new learning mechanisms to The rapid growth of social media sites created a wide
analyze interactions of a user with the system and its ability opportunity to build social recommender systems. The
to convert it into user preference can make recommender clustering of users, according to their tastes as a similar
system more dynamic in providing suggestions. As a hybrid metric, can generate good recommendation in more efficient
approach utilization of ontologies may be used to represent manner. Investigation of problems such as influence of
the user’s preferences in the semantic manner, such approach friends and their distance, time-series information, acqui-
can overcome difficulties in the lack of personalization with sition of tastes of individual user, enabling of privacy and
the textual information. The location information is already security of the users and their data, dynamic variations and
used by many recommender systems, which can be followed places of the user, automated analysis of heterogeneous data
by utilization of device sensors’ data such as RFID signals, through a flexible framework, practical situation influenced
weather temperature, and health metrics/signals. Initially, improvement of performance, and cold start problem may be
Computational Intelligence and Neuroscience 23

1.0 1.0

0.8 0.8

0.6 0.6
NDCG

NDCG
0.4 0.4

0.2 0.2

0.0 0.0
2 3 4 5 6 7 8 2 3 4 5 6 7 8
Group of users (size) Group of users (size)
NDCG@5 NDCG@5
NDCG@10 NDCG@10
NDCG@20 NDCG@20
(a) (b)
1.0 1.6

0.8
1.2

0.6
NDCG

MAE

0.8
0.4

0.4
0.2

0.0 0.0
2 3 4 5 6 7 8 2 3 4 5 6 7 8
Group of users (size) Group of users (size)
NDCG@5 Similar users
NDCG@10 Dissimilar users
NDCG@20 Random users
(c) (d)

Figure 14: (a) Comparison of NDCG for highly similar users. (b) Comparison of NDCG for dissimilar users. (c) Comparison of NDCG for
random users. (d) Comparison of MAE for highly similar, dissimilar, and random users.

considered in the discovery and development of new hybrid algorithm more reliable and helps to make recommendations
recommender systems. As a crucial conclusion, the success more quickly through the improved computational efficiency.
of recommender systems purely depends on the effective The SPTW based location recommendation system is
learning of user behavior and generation of user acceptable extended for the group of users. The system is evaluated for
recommendations. group of users recommendations through the groups formed
Location recommendation system is proposed and eval- in foursquare dataset. The group size ranges from 2 to 8
uated using foursquare dataset. The ratings are predicted and the groups are classified as highly similar, random, and
through social pertinent trust walker algorithm. The results dissimilar. The groups are classified on the basis of uniformity
are compared with the other existing TrustWalker methods, between users to maintain cohesiveness and bondage. The
namely, TidalTrust, MoleTrust, TrustWalker, and Relevant- results are evaluated based on NDCG and MAE metrics. The
TrustWalker. It is observed from the experimental results that inference on the results of evaluation process shows that the
RMSE, coverage, precision, and 𝐹-measure for the proposed efficiency and performance of the model reduce as the size
SPTW based location recommender system achieve the best of the group increases. The ultimate outcome of the results
performance. Compared to other models that consider trust shows that the groups with more similar users perform well
between the users, the SPTW costs less time because, in every compared to random users and dissimilar users. The average
iteration or walk, SPTW chooses target node based on proba- processing time taken for the various types of groups also
bility and trust relevancy. Such an approach makes the SPTW supports the performance inference of the recommendation
24 Computational Intelligence and Neuroscience

450 600

367.9
456.28
450
304.8 382.49
300

Processing time
Processing time

269.4 351.64

208.6 219.5 300

201.3
180.2
207.32
150 168.85
150 142.47

0 0
2 3 4 5 6 7 8 Similar Dissimilar Random Male Female All users
Group of users (size) Types of users
(a) (b)

Figure 15: (a) Comparison of processing time of SPTW-GRM for various group sizes. (b) Comparison of average processing time of SPTW-
GRM for different users.

results. SPTW-GRM proves its efficiency and performance the strength of web based applications with their social
on both recommendations for individual users and group of network exposure. Through offering social tools, web apps
users. promote the use of filtering mechanisms such as collaborative
filtering. It is well known that these collaborative filtering
9.1. Future Work Guidelines. In this section, we quickly techniques allow rating of items and collecting data of the
remark the important issues that are examined at present in user in the social level or individual level. The systems which
the advancement of recommender system frameworks in e- use such mechanisms are called social recommender systems.
Tourism domain, such as These available social tools can be used to differentiate the
cluster of users from cluster of items. In an existing work,
(i) broadening of the recommendations offered to the the users hold a related tag cloud pertinent to their profile
client; and another tag is created for their interests. Such a kind of
(ii) utilization of social information accessible in the data is used to discover the coincidence between the users and
present Web 2.0 apps; items. Some travel recommenders grab the data regarding the
(iii) change in improved recommendations through uti- connections of the users along with the analysis of searches
lizing the additional capacities of smart mobile and readings in the wiki. These recommender systems utilize
phones. this data to calculate the satisfaction degree of the user for
the particular article. Another existing recommender system
What Is the Need for Content-Based Systems in Tourism? keeps up the social media profile of the user to consider
The main concentration of the content-based systems is to the contact information and to examine the communication
recommend things to the users based on their profile, by between other users. The main concept while considering the
generating the user specific results which may fascinate the social recommender systems is trust. Trust is very important
user. In the area of tourism, this may be considered as an challenge as it deals with the ratings of the items and
essential issue of travel recommender systems. Few recom- the reliability of the user. While addressing the issues in
mender systems try to promote new spots or new activities, the segment of trust in social recommender systems, the
which makes the recommendations ineffective. A good travel significant aspect to be considered is that the user with
recommender system provides broad suggestions to the users higher reliability score should be treated with higher weights
and allows them to choose their routes with activities. The compared to others with the lower weights.
list of user’s attractions should be synchronized with rating
limit in order to maintain the quality of recommendations. Why Adaptability Is Considered as Very Important Quality of
The usage of multiple techniques to filter the activities for the Travel Recommender Systems? A unique feature of tourism
recommendation generation is a new scope in this domain. domain is the area where the recommenders have been
The clustering of items into groups with similar attributes is used, as it adapts to the users and helps them through
an effective mechanism in the recommendation process to generated suggestions in different places and in different
build a good list of suggestions by utilizing every cluster of moments. These travel recommender systems have begun to
items for a user specific proposition of their tourist interests. fuse context aware mechanisms with it. The accomplishment
of this methodology is because of the far reaching utilization
Is Growth of Web Treated as an Issue with Travel Recommender of Smartphones. Numerous tourism recommenders keep
Systems? In the present scenario, recommender systems use running mobile devices, so the location of the user may be
Computational Intelligence and Neuroscience 25

utilized in the filtering process of items to be demonstrated. In Hypermedia and Adaptive Web-Based Systems (AH ’02), Malaga,
contrast to the existing systems, the enhancement of location Spain, 2002.
consideration has to be modified. The present location of [4] K.-W. Cheung, J. T. Kwok, M. H. Law, and K.-C. Tsui, “Mining
the user is important. But along with it, the significance of customer product ratings for personalized marketing,” Decision
other spots visited already should also be analyzed and used Support Systems, vol. 35, no. 2, pp. 231–243, 2003.
in the recommendations. In tourism recommender systems, [5] L. Stanley, F. Lorenzi, R. Saldaña, and D. Licthnow, “A tourism
different highlights are considered as related data for an recommender system based on collaboration and text analysis,”
event. For example, the present climate is analyzed to choose Information Technology and Tourism, vol. 6, pp. 157–165, 2003.
appropriate outdoor or indoor activities to be recommended. [6] T. Berners-Lee, J. Hendler, and O. Lassila, “The semantic web,”
Scientific American Magazine, vol. 284, pp. 34–43, 2001.
What Is the Requirement of Hybrid Recommendation Models [7] P. Li and S. Yamada, “A movie recommender system based on
in TRS? In a travel recommendation framework, a mul- inductive learning,” in Proceedings of the IEEE Conference on
tifarious model of the context is considered for building Cybernetics and Intelligent Systems, vol. 1, pp. 318–323, IEEE,
customized route arrangements. The extracted context data December 2004.
is sorted out in an order, with features identified with the cli- [8] J. Shi, J. Chen, and Z. Bao, “An application study on collaborative
mate, travel traffic, security (such as parking, mobile connec- filtering in e-commerce,” in Proceedings of the 8th Interna-
tivity, and medicinal facilities), service facilities (vehicle ser- tional Conference on Service Systems and Service Management
vice station, etc.), and vacation spots (beach, entertainment (ICSSSM ’11), Tianjin, China, June 2011.
place, fishing zone, etc.). The generally defined four prime [9] W. Carrer-Neto, M. L. Hernández-Alcaraz, R. Valencia-Garcı́a,
parameters of context are location (e.g., current location of and F. Garcı́a-Sánchez, “Social knowledge-based recommender
the of the user and the locality of spot), time (time required system. Application to the movies domain,” Expert Systems with
by the client to achieve the spot, the opening/shutting Applications, vol. 39, no. 12, pp. 10990–11000, 2012.
times, etc.), climate and natural conditions (e.g., temperature, [10] P. Winoto and T. Y. Tang, “The role of user mood in movie
mugginess, precipitation degree, wind, season, snippet of the recommendations,” Expert Systems with Applications, vol. 37, no.
day), and social elements (number of clients near to the 8, pp. 6086–6092, 2010.
spot and number of positive/negative inputs). These aspects [11] S. K. Lee, Y. H. Cho, and S. H. Kim, “Collaborative filtering
were formulated as key points in the accomplishment of with ordinal scale-based implicit ratings for mobile music
travel recommender systems in the area of e-Tourism, due recommendations,” Information Sciences, vol. 180, no. 11, pp.
to the intrinsic mobile activity of the users in this particular 2142–2155, 2010.
application area. [12] A. Nanolopoulus, D. Rafailidis, P. Symeonidis, and Y. Manolo-
poulus, “MusicBox: personalized music recommendation based
on cubic analysis of social tags,” IEEE Transactions on Audio,
Competing Interests Speech, and Language Processing, vol. 18, no. 2, pp. 407–412, 2010.
[13] E. R. Núñez-Valdéz, J. M. Cueva Lovelle, O. Sanjuán Martı́nez,
The authors declare that there are no competing interests V. Garcı́a-Dı́az, P. Ordoñez De Pablos, and C. E. Montenegro
regarding the publication of this paper. Marı́n, “Implicit feedback techniques on recommender systems
applied to electronic books,” Computers in Human Behavior, vol.
28, no. 4, pp. 1186–1193, 2012.
Acknowledgments [14] Z. Huang, D. Zeng, and H. Chen, “A comparison of collabora-
The authors sincerely thank DST-SERB (YSS/2014/000718) tive-filtering algorithms for ecommerce,” IEEE Intelligent Sys-
and SASTRA University for providing High Performance tems, vol. 22, no. 5, pp. 68–78, 2007.
Computing Cluster (HPCC) facility and great support that [15] J. J. Castro-Schez, R. Miguel, D. Vallejo, and L. M. López-López,
enabled us to carry out this research work. “A highly adaptive recommender system based on fuzzy logic
for B2C e-commerce portals,” Expert Systems with Applications,
vol. 38, no. 3, pp. 2441–2454, 2011.
References [16] D. Rosaci and G. M. L. Sarné, “A multi-agent recommender sys-
tem for supporting device adaptivity in e-Commerce,” Journal of
[1] P. Resnick, N. Iakovou, M. Sushak, P. Bergstrom, and J. Riedl, Intelligent Information Systems, vol. 38, no. 2, pp. 393–418, 2012.
“GroupLens: an open architecture for collaborative filtering of
[17] O. Zaiane, “Building a recommender agent for e-learning sys-
netnews,” in Proceedings of the ACM Conference on Computer
tems,” in Proceedings of the International Conference on Comput-
Supported Cooperative Work (CSCW ’94), pp. 175–186, Chapel
ers in Education (ICCE ’02), vol. 1, pp. 55–59, IEEE, Auckland,
Hill, NC, USA, 1994.
New Zealand, December 2002.
[2] U. Shardanand and P. Maes, “Social information filtering: [18] J. Bobadilla, F. Serradilla, and A. Hernando, “Collaborative fil-
algorithms for automating ‘Word of Mouth’,” in Proceedings of tering adapted to recommender systems of e-learning,” Knowl-
the Human Factors in Computing Systems Conference, pp. 210– edge-Based Systems, vol. 22, no. 4, pp. 261–265, 2009.
217, May 1995.
[19] E. Costa-Montenegro, A. B. Barragáns-Martı́nez, and M. Rey-
[3] R. Ghani and A. Fano, “Building recommender systems using López, “Which App? A recommender system of applications in
a knowledge base of product semantics,” in Proceedings of markets: implementation of the service for monitoring users’
the Workshop on Recommendation and Personalization in interaction,” Expert Systems with Applications, vol. 39, no. 10, pp.
eCommerce at the 2nd International Conference on Adaptive 9367–9375, 2012.
26 Computational Intelligence and Neuroscience

[20] F. Ricci, “Travel recommender systems,” IEEE Intelligent Sys- [37] J. Bobadilla, F. Ortega, and A. Hernando, “A collaborative
tems, vol. 17, pp. 55–57, 2002. filtering similarity measure based on singularities,” Information
[21] M. Gao, K. Liu, and Z. Wu, “Personalisation in web computing Processing and Management, vol. 48, no. 2, pp. 204–217, 2012.
and informatics: theories, techniques, applications, and future [38] B. Jeong, J. Lee, and H. Cho, “User credit based collaborative
research,” Information Systems Frontiers, vol. 12, no. 5, pp. 607– filtering,” Expert Systems with Applications, vol. 36, no. 3, pp.
629, 2010. 7309–7312, 2009.
[22] G. Adomavicius and A. Tuzhilin, “Toward the next generation [39] H.-N. Kim, A. El-Saddik, and G.-S. Jo, “Collaborative error-
of recommender systems: a survey of the state-of-the-art and reflected models for cold-start recommender systems,” Decision
possible extensions,” IEEE Transactions on Knowledge and Data Support Systems, vol. 51, no. 3, pp. 519–531, 2011.
Engineering, vol. 17, no. 6, pp. 734–749, 2005. [40] J. Borràs, A. Moreno, and A. Valls, “Intelligent tourism recom-
[23] J. Ben Schafer, D. Frankowski, J. Herlocker, and S. Sen, “Col- mender systems: a survey,” Expert Systems with Applications,
laborative filltering recommender systems,” in The Adaptive vol. 41, no. 16, pp. 7370–7389, 2014.
Web, P. Brusilovsky, A. Kobsa, and W. Nejdl, Eds., vol. 4321 of [41] Y. Kurata, “CT-planner2: more flexible and interactive assis-
Lecture Notes in Computer Science, pp. 291–324, Springer, Berlin, tance for day tour planning,” in Proceedings of the International
Germany, 2007. Conference on Information and Communication Technologies in
Tourism (ENTER ’11), pp. 25–37, Innsbruck, Austria, January
[24] A. M. Rashid, G. Karypis, and J. Riedl, “Learning preferences of
2011.
new users in recommender systems: an information theoretic
approach,” ACM SIGKDD Explorations Newsletter, vol. 10, no. [42] S. Venkataiaha, N. Shardaa, and M. Ponnadaa, “A comparative
2, pp. 90–100, 2008. study of continuous and discrete visualisation of tourism
information,” in Information and Communication Technologies
[25] P. du Boucher-Ryan and D. Bridge, “Collaborative recommend- in Tourism 2008: Proceedings of the International Conference in
ing using formal concept analysis,” Knowledge-Based Systems, Innsbruck, Austria, 2008, pp. 12–23, Springer, Berlin, Germany,
vol. 19, no. 5, pp. 309–315, 2006. 2008.
[26] Y.-J. Park and A. Tuzhilin, “The long tail of recommender [43] C.-S. Lee, Y.-C. Chang, and M.-H. Wang, “Ontological recom-
systems and how to leverage it,” in Proceedings of the 2nd ACM mendation multi-agent for Tainan city travel,” Expert Systems
International Conference on Recommender Systems (RecSys ’08), with Applications, vol. 36, no. 3, pp. 6740–6753, 2009.
pp. 11–18, ACM, Lousanne, Switzerland, October 2008.
[44] P. Vansteenwegen, W. Souffriau, G. Vanden Berghe, and D.
[27] S.-T. Park and W. Chu, “Pairwise preference regression for Van Oudheusden, “The city trip planner: an expert system for
cold-start recommendation,” in Proceedings of the 3rd ACM tourists,” Expert Systems with Applications, vol. 38, no. 6, pp.
Conference on Recommender Systems (RecSys ’09), pp. 21–28, 6540–6546, 2011.
ACM, Boston, Mass, USA, October 2009. [45] L. Sebastià, I. Garcia, E. Onaindia, and C. Guzman, “E-tourism:
[28] D. Jannach, “Fast computation of query relaxations for knowl- a tourist recommendation and planning application,” Inter-
edge-based recommenders,” AI Communications, vol. 22, no. 4, national Journal on Artificial Intelligence Tools, vol. 18, no. 5, pp.
pp. 235–248, 2009. 717–738, 2009.
[29] X. Luo, Y. Xia, and Q. Zhu, “Incremental Collaborative Filtering [46] A. Montejo-Ráez, J. M. Perea-Ortega, M. Á. Garcı́a-Cumbreras,
recommender based on Regularized Matrix Factorization,” and F. Martı́nez-Santiago, “Otiûm: a web based planner for
Knowledge-Based Systems, vol. 27, pp. 271–280, 2012. tourism and leisure,” Expert Systems with Applications, vol. 38,
[30] J. Bobadilla and F. Serradilla, “The effect of sparsity on collabo- no. 8, pp. 10085–10093, 2011.
rative filtering metrics,” in Proceedings of the 20th Australasian [47] J. Borràs, A. Moreno, A. Valls et al., “Uso de técnicas de
Database Conference (ADC ’09), pp. 9–17, January 2009. inteligencia artificial para hacer recomendaciones enoturı́sticas
[31] J. Bobadilla, A. Hernando, F. Ortega, and J. Bernal, “A frame- personalizadas en la Provincia de Tarragona,” in Proceedings of
work for collaborative filtering recommender systems,” Expert the 9th Congreso Turismo y Tecnologı́as de la Información y las
Systems with Applications, vol. 38, no. 12, pp. 14609–14623, 2011. Comunicaciones, pp. 217–230, Málaga, Spain, 2012.
[48] D. Jannach, M. Zanker, and M. Jessenitschnig, “Developing
[32] B. Sarwar, G. Karypis, J. A. Konstan, and J. Reidl, “Item-
knowledge-based travel advisor systems: a case study,” in
based collaborative filtering recommendation algorithms,” in
Tourism Informatics: Visual Travel Recommender Systems, Social
Proceedings of the 10th International Conference on World Wide
Communities, and User Interface Design, N. Sharda, Ed., pp. 38–
Web, pp. 285–295, Hong Kong, May 2001.
53, Information Science Reference, Hershey, Pa, USA, 2010.
[33] M. Gao, Z. Wu, and F. Jiang, “Userrank for item-based col- [49] W. Wang, G. Zeng, and D. Tang, “Bayesian intelligent semantic
laborative filtering recommendation,” Information Processing mashup for tourism,” Concurrency and Computation: Practice
Letters, vol. 111, no. 9, pp. 440–446, 2011. and Experience, vol. 23, no. 8, pp. 850–862, 2011.
[34] W. Qin, L. Xin, and H. Liang, “Unifying user-based and item- [50] C.-C. Yu and H. Ping Chang, “Personalized location-based
based algorithm to improve collaborative filtering accuracy,” recommendation services for tour planning in mobile tourism
Energy Procedia, vol. 13, pp. 8231–8239, 2011. applications,” in EC-Web, T. D. Noia and F. Buccafurri, Eds., pp.
[35] L. Candillier, F. Meyer, and M. Boullé, “Comparing state-of- 38–49, Springer, 2009.
the-art collaborative filtering systems,” in Machine Learning and [51] D. Gavalas and M. Kenteris, “A web-based pervasive recom-
Data Mining in Pattern Recognition, P. Perner, Ed., vol. 4571 of mendation system for mobile tourist guides,” Personal and
Lecture Notes in Computer Science, pp. 548–562, Springer, 2007. Ubiquitous Computing, vol. 15, no. 7, pp. 759–770, 2011.
[36] J. Bobadilla, F. Serradilla, and J. Bernal, “A new collaborative [52] F. Ricci, Q. N. Nguyen, and O. Averjanova, “Exploiting a map-
filtering metric that improves the behavior of recommender based interface in conversational recommender systems for
systems,” Knowledge-Based Systems, vol. 23, no. 6, pp. 520–528, mobile travelers,” in Tourism Informatics: Visual Travel Recom-
2010. mender Systems, Social Communities, and User Interface Design,
Computational Intelligence and Neuroscience 27

N. Sharda, Ed., pp. 73–93, Information Science Reference, [66] A. Luberg, T. Tammet, and P. Järv, “Smart city: a rule-based
Hershey, Pa, USA, 2010. tourist recommendation system,” in Proceedings of the Informa-
[53] F. Martı́nez-Santiago, F. Ariza-López, A. Montejo-Ráez, and tion and Communication Technologies in Tourism (ENTER ’11),
A. Ureña-López, “GeOasis: a knowledge-based geo-referenced R. Law, M. Fuchs, and F. Ricci, Eds., pp. 51–62, 2011.
tourist assistant,” Expert Systems with Applications, vol. 39, no. [67] M. Batet, A. Moreno, D. Sánchez, D. Isern, and A. Valls,
14, pp. 11737–11745, 2012. “Turist@: agent-based personalised recommendation of tourist
activities,” Expert Systems with Applications, vol. 39, no. 8, pp.
[54] M. Rey-López, A. B. Barragáns-Martı́nez, A. Peleteiro, F. 7319–7329, 2012.
A. Mikic-Fonte, and J. C. Burguillo, “moreTourism: mobile
[68] J. Borràs, A. Valls, A. Moreno, and D. Isern, “Ontology-based
recommendations for tourism,” in Proceedings of the IEEE
management of uncertain preferences in user profiles,” in
International Conference on Consumer Electronics (ICCE ’11), pp.
Advances in Computational Intelligence, S. Greco, B. Bouchon-
347–348, Las Vegas, Nev, USA, January 2011.
Menier, B. Bouchon-Menier et al., Eds., vol. 298 of Commu-
[55] D. Martin, A. Alzua, and C. Lamsfus, “A contextual geofencing nications in Computer and Information Science, pp. 127–136,
mobile tourism service,” in Proceedings of the International Springer, Berlin, Germany, 2012.
Conference on Information and Communication Technologies in [69] A. Garcia, P. Vansteenwegen, O. Arbelaitz, W. Souffriau, and
Tourism (ENTER ’11), R. Law, M. Fuchs, and F. Ricci, Eds., pp. M. T. Linaza, “Integrating public transportation in personalised
191–202, Innsbruck, Austria, January 2011. electronic tourist guides,” Computers & Operations Research,
[56] M. Braunhofer, M. Elahi, F. Ricci, and T. Schievenin, “Context- vol. 40, no. 3, pp. 758–774, 2013.
aware points of interest suggestion with dynamic weather data [70] R. Gyorodi, C. Gyorodi, and M. Dersidan, “An extended rec-
management,” in Information and Communication Technologies ommendation system using data mining implemented for smart
in Tourism 2014, pp. 87–100, Springer, 2014. phones,” International Journal of Computers and Technology, vol.
[57] A. Umanets, A. Ferreira, and N. Leite, “GUIDEME—a tourist 11, no. 3, pp. 2360–2372, 2013.
guide with a recommender system and social interaction,” [71] Y. Kurata and T. Hara, “CT-planner4: toward a more user-
Procedia Technology, vol. 17, pp. 407–414, 2014, Proceedings of the friendly interactive day-tour planner,” in Information and Com-
Conference on Electronics, Telecommunications and Computers munication Technologies in Tourism 2014: Proceedings of the
(CETC ’13). International Conference in Dublin, Ireland, January 21–24, 2014,
pp. 73–86, Springer, Berlin, Germany, 2014.
[58] J. M. Noguera, M. J. Barranco, R. J. Segura, and L. Martı́nez,
[72] J. P. Lucas, N. Luz, M. N. Moreno, R. Anacleto, A. A. Figueiredo,
“A mobile 3D-GIS hybrid recommender system for tourism,”
and C. Martins, “A hybrid recommendation approach for a
Information Sciences, vol. 215, pp. 37–52, 2012.
tourism system,” Expert Systems with Applications, vol. 40, no.
[59] F. Lorenzi, S. Loh, and M. Abel, “PersonalTour: a recom- 9, pp. 3532–3550, 2013.
mender system for travel packages,” in Proceedings of the [73] K. Meehan, T. Lunney, K. Curran, and A. McCaughey, “Context-
IEEE/WIC/ACM International Conference on Intelligent Agent aware intelligent recommendation system for tourism,” in
Technology (IAT ’11), pp. 333–336, Lyon, France, August 2011. Proceedings of the IEEE International Conference on Pervasive
[60] I. Seidel, M. Gärtner, M. Pöttler, H. Berger, and M. Dittenbach, Computing and Communications Workshops (PerCom ’13), pp.
“Itchy feet: a 3D e-tourism environment,” in Tourism Informat- 328–331, San Diego, Calif, USA, March 2013.
ics: Visual Travel Recommender Systems, Social Communities, [74] T. Ruotsalo, K. Haav, A. Stoyanov et al., “SMARTMUSEUM:
and User Interface Design, N. Sharda, Ed., IGI Global, Hershey, a mobile recommender system for the web of data,” Web
Pa, USA, 2009. Semantics: Science, Services and Agents on the World Wide Web,
[61] S. Koceski and B. Petrevska, “Empirical evidence of contri- vol. 20, pp. 50–67, 2013.
bution to e-tourism by application of personalized tourism [75] A. Savir, R. I. Brafman, and G. Shani, “Recommending
recommendation system,” Annals of the Alexandru Ioan Cuza improved configurations for complex objects with an appli-
University—Economics, vol. 59, no. 1, pp. 363–374, 2013. cation in travel planning,” in Proceedings of the 7th ACM
Conference on Recommender Systems (RecSys ’13), pp. 391–394,
[62] I. Mı́nguez, D. Berrueta, and L. Polo, “CRUZAR: an application
Hong Kong, October 2013.
of semantic matchmaking to e-tourism,” in Cases on Semantic
Interoperability for Information Systems Integration: Practices [76] W.-S. Yang and S.-Y. Hwang, “ITravel: a recommender system in
and Applications, Y. Kalfoglou, Ed., pp. 255–271, Information mobile peer-to-peer environment,” The Journal of Systems and
Science Reference, Hershey, Pa, USA, 2010. Software, vol. 86, no. 1, pp. 12–20, 2013.
[77] P. Vansteenwegen and W. Souffriau, “Trip planning function-
[63] L. Sebastia, A. Giret, and I. Garcia, “A multi agent archi- alities: state of the art and future,” Information Technology &
tecture for tourism recommendation,” in Trends in Practical Tourism, vol. 12, no. 4, pp. 305–315, 2010.
Applications of Agents and Multiagent Systems, Y. Demazeau,
[78] L. Castillo, E. Armengol, E. Onaindı́a et al., “Samap: an user-
F. Dignum, J. M. Corchado et al., Eds., vol. 71 of Advances in
oriented adaptive system for planning tourist visits,” Expert
Intelligent and Soft Computing, pp. 547–554, Springer, Berlin,
Systems with Applications, vol. 34, no. 2, pp. 1318–1332, 2008.
Germany, 2010.
[79] L. Ceccaroni, V. Codina, M. Palau, and M. Pous, “PaTac: urban,
[64] J. Borràs, J. de la Flor, Y. Pérez et al., “SigTur/E-destination: ubiquitous, personalized services for citizens and tourists,” in
a system for the management of complex tourist regions,” in Proceedings of the3rd International Conference on Digital Society
Information and Communication Technologies in Tourism 2011, (ICDS ’09), pp. 7–12, IEEE, Cancun, Mexico, February 2009.
R. Law, M. Fuchs, and F. Ricci, Eds., pp. 39–50, Springer, 2011. [80] Y. Huang and L. Bian, “A Bayesian network and analytic hier-
[65] I. Garcia, L. Sebastià, and E. Onaindia, “On the design of indi- archy process based personalized recommendations for tourist
vidual and group recommender systems for tourism,” Expert attractions over the Internet,” Expert Systems with Applications,
Systems with Applications, vol. 38, no. 6, pp. 7683–7692, 2011. vol. 36, no. 1, pp. 933–943, 2009.
28 Computational Intelligence and Neuroscience

[81] V. Subramaniyaswamy and S. C. Pandian, “Effective tag recom- [98] D. Sánchez and A. Moreno, “Learning non-taxonomic relation-
mendation system based on topic ontology using wikipedia and ships from web documents for domain ontology construction,”
WordNet,” International Journal of Intelligent Systems, vol. 27, Data and Knowledge Engineering, vol. 64, no. 3, pp. 600–623,
no. 12, pp. 1034–1048, 2012. 2008.
[82] V. Subramaniyaswamy, V. Vijayakumar, and V. Indragandhi, [99] D. Sánchez, A. Moreno, and L. Del Vasto-Terrientes, “Learning
“A review of ontology-based tag recommendation approaches,” relation axioms from text: an automatic web-based approach,”
International Journal of Intelligent Systems, vol. 28, no. 11, pp. Expert Systems with Applications, vol. 39, no. 5, pp. 5792–5805,
1054–1071, 2013. 2012.
[83] V. Subramaniyaswamy, V. Vijayakumar, V. Indragandhi, and R. [100] J. M. Ruiz-Martı́nez, J. A. Miñarro-Giménez, D. Castellanos-
Logesh, “Data mining-based tag recommendation system: an Nieves, F. Garcı́a-Sáanchez, and R. Valencia-Garcı́a, “Ontol-
overview,” WIREs Data Mining and Knowledge Discovery, vol. ogy population: an application for the e-tourism domain,”
5, no. 3, pp. 87–112, 2015. International Journal of Innovative Computing, Information and
[84] I. Garcı́a-Manotas, E. Lupiani, F. Garcia-Sánchez, and R. Control, vol. 7, no. 11, pp. 6115–6134, 2011.
Valencia-Garcı́a, “Populating knowledge based decision sup- [101] J. Bao, Y. Zheng, D. Wilkie, and M. Mokbel, “Recommendations
port systems,” International Journal of Decision Support Systems in location-based social networks: a survey,” GeoInformatica,
Technology, vol. 2, pp. 1–20, 2010. vol. 19, no. 3, pp. 525–565, 2015.
[85] G. Fenza, E. Fischetti, D. Furno, and V. Loia, “A hybrid con- [102] Y. Zheng, “Location-based social networks: users,” in Comput-
text aware system for tourist guidance based on collaborative ing with Spatial Trajectories, Y. Zheng and X. Zhou, Eds., pp.
filtering,” in Proceedings of the IEEE International Conference on 243–276, Springer, Berlin, Germany, 2011.
Fuzzy Systems (FUZZ ’11), pp. 131–138, Taipei, Taiwan, June 2011.
[103] X. Xiao, Y. Zheng, Q. Luo, and X. Xie, “Finding similar users
[86] F.-M. Hsu, Y.-T. Lin, and T.-K. Ho, “Design and implementation using category-based location history,” in Proceedings of the
of an intelligent recommendation system for tourist attractions: 18th ACM SIGSPATIAL International Conference on Advances in
the integration of EBM model, Bayesian network and Google Geographic Information Systems, pp. 442–445, ACM, San Jose,
Maps,” Expert Systems with Applications, vol. 39, no. 3, pp. 3257– Calif, USA, November 2010.
3264, 2012.
[104] N. Eagle and A. Pentland, “Reality mining: sensing complex
[87] A. Moreno, A. Valls, D. Isern, L. Marin, and J. Borràs, “SigTur/e- social systems,” Personal and Ubiquitous Computing, vol. 10, no.
destination: ontology-based personalized recommendation of 4, pp. 255–268, 2006.
tourism and leisure activities,” Engineering Applications of Arti-
[105] M. Jamali and M. Ester, “TrustWalker: a random walk model for
ficial Intelligence, vol. 26, no. 1, pp. 633–651, 2013.
combining trust-based and item-based recommendation,” in
[88] M. Wooldridge, An Introduction to Multiagent Systems, John
Proceedings of the 15th ACM SIGKDD International Conference
Wiley and Sons, 2009.
on Knowledge Discovery and Data Mining (KDD ’09), pp. 397–
[89] W. Souffriau, P. Vansteenwegen, G. Vanden Berghe, and D. Van 406, Paris, France, July 2009.
Oudheusden, “The planning of cycle trips in the province of
[106] A. Jøsang, R. Ismail, and C. Boyd, “A survey of trust and rep-
East Flanders,” Omega, vol. 39, no. 2, pp. 209–213, 2011.
utation systems for online service provision,” Decision Support
[90] B. Dasarathy, Nearest Neighbour (NN) Norms: NN Pattern Clas- Systems, vol. 43, no. 2, pp. 618–644, 2007.
sification Techniques, IEEE Computer Society Press, 1991.
[107] S. Berkovsky, J. Freyne, and M. Coombe, “Aggregation trade
[91] L. Martı́nez, R. M. Rodrı́guez, and M. Espinilla, “Reja: a
offs in family based recommendations,” in AI 2009: Advances
georeferenced hybrid recommender system for restaurants,” in
in Artificial Intelligence: 22nd Australasian Joint Conference,
Proceedings of the IEEE/WIC/ACM International Conference on
Melbourne, Australia, December 1–4, 2009. Proceedings, A. E.
Web Intelligence and Intelligent Agent Technology (WI-IAT ’09),
Nicholson and X. Li, Eds., vol. 5866 of Lecture Notes in
pp. 187–190, Milan, Italy, September 2009.
Computer Science, pp. 646–655, Springer, Berlin, Germany,
[92] J. J. Dujmović and H. Nagashima, “LSP method and its use for 2009.
evaluation of Java IDEs,” International Journal of Approximate
[108] M. S. Pera and Y.-K. Ng, “A group recommender for movies
Reasoning, vol. 41, no. 1, pp. 3–22, 2006.
based on content similarity and popularity,” Information Pro-
[93] R. R. Yager, “On ordered weighted averaging aggregation cessing and Management, vol. 49, no. 3, pp. 673–687, 2013.
operators in multicriteria decisionmaking,” IEEE Transactions
[109] J. Lee, “Analyses of multiple evidence combination,” in Proceed-
on Systems, Man and Cybernetics, vol. 18, no. 1, pp. 183–190, 1988.
ings of the 20th Annual International ACM SIGIR Conference on
[94] A. Valls, “ClusDM: a multiple criteria decision method for
Research and Development in Information Retrieval (SIGIR ’97),
heterogeneous data sets (Ph.D. thesis),” AI Communications,
pp. 267–276, Philadelphia, Pa, USA, July 1997.
vol. 16, no. 2, pp. 129–130, 2003.
[110] J. A. Golbeck, Computing and applying trust in web-based social
[95] A. Garcı́a-Crespo, J. Chamizo, I. Rivera, M. Mencke, R.
networks [Ph.D. thesis], University of Maryland, 2005.
Colomo-Palacios, and J. M. Gómez-Berbı́s, “SPETA: social
pervasive e-tourism advisor,” Telematics and Informatics, vol. [111] P. Massa and P. Avesani, “Trust-aware recommender systems,”
26, no. 3, pp. 306–315, 2009. in Proceedings of the ACM Conference on Recommender Systems
(RecSys ’07), pp. 17–24, Minneapolis, Minn, USA, October 2007.
[96] C. Lamsfus, A. Alzua-Sorzabal, D. Martin, Z. Salvador, and
A. Usandizaga, “Human-centric ontology-based context mod- [112] S. Deng, L. Huang, and G. Xu, “Social network-based service
elling in tourism,” in Proceedings of the International Conference recommendation with trust enhancement,” Expert Systems with
on Knowledge Engineering and Ontology Development (KEOD Applications, vol. 41, no. 18, pp. 8075–8084, 2014.
’09), pp. 424–434, Madeira, Portugal, October 2009. [113] S. Amer-Yahia, S. Basu Roy, A. Chawlat, G. Das, and C. Yu,
[97] D. Sánchez and A. Moreno, “Pattern-based automatic taxonomy “Group recommendation: semantics and efficiency,” Very Large
learning from the web,” AI Communications, vol. 21, no. 1, pp. Data Base Endowment, vol. 2, pp. 754–765, 2009.
27–48, 2008.