Context-aware and Ontology-based Recommender System for e-Tourism

Gustavo Castellanos

, Yudith Cardinale

1,2 a

and Philippe Roose

Computer Science Department, Universidad Sim

on Bol

ıvar, Caracas 1080, Venezuela

School of Electronics and Telecommunications Engineering, Universidad Cat

olica San Pablo, Arequipa, 04001, Peru

Universit

e de Pau et des Pays de l’Adour, LIUPPA – T2I64600, Anglet, France

Keywords:

Tourism, Recommender Systems, Context Awareness, Ontology, Serendipity, User-centric.

Abstract:

Frequently, travelers try to collect information for planing a trip or when being at the destination. Usually,

tourists depend on places’ reviews to make the choice, but this implies prior knowledge of the touristic places

and explicit search for suggestions through interaction with applications (i.e., PULL paradigm). In contrast, a

PUSH approach, in which the application proactively triggers a recommendation process according to users’

preferences and when necessary, seems to be a more reasonable solution. Recommender systems have be-

come appropriate applications to help tourists in their trip planning. However, they still have limitations, such

as poor consideration of users’ proﬁles and their contexts, their predictable suggestions, and the lack of a

standard representation of the knowledge managed. We propose a user-centric recommender system archi-

tecture, that supports both PULL and PUSH approaches, assisted by an ontology-based spreading activation

algorithm for context-aware recommendations, with a focus on decreasing predictable outputs and increasing

serendipity, based on an aging-like approach. To demonstrate its suitability and performance, we develop a

ﬁrst prototype of the architecture and simulate different scenarios, varying users’ proﬁles, preferences, and

context parameters. Results show that the ontology-based spreading activation and the proposed aging system

provide relevant and varied recommendations according to users’ preferences, while considering their context

and improving the serendipity of the system when comparing with a state-of-the-art work.

1 INTRODUCTION

Tourism is one of the most promising areas for sup-

porting the economy of countries and is becoming an

extremely important market (Alghamdi et al., 2016;

Artemenko et al., 2017; Kazandzhieva and Santana,

2019; Kayumovich, 2020). When travellers plan to go

to a destination, they try to collect information about

their new destination “as best as they can” (e.g., by

going to the tourist ofﬁce or by obtaining informa-

tion about their destination and its surroundings on

Internet). This task may be overwhelming due to the

big amount of available information sources. Usually,

tourists depend on places’ reviews to decide where to

spend their vacations, their free time, or even to take

lunch, but this requires prior knowledge of the exis-

tence of the places.

Currently, there is a plethora of applications offer-

ing information of historical centres, points of interest

(POI), city tours, green areas to rest, etc. The use of

such applications implies that the user is ﬁrstly aware

https://orcid.org/0000-0002-5966-0113

of them; then installs them in the hope that they are

suitable (an average of 3.3% of mobile apps are still

being used after 30 days of being installed

, and from

an average of 40 installed apps, the 89% of the time

of use is dedicated to 18 of such apps

Most of these applications follow the PULL

paradigm, i.e., the user has to explicitly search

for suggestions/recommendations through interaction

with the application’s graphical user interface. We be-

lieve that this is one of the reasons why these types of

applications are not successful: they require too many

steps, too many interactions, and too many unknowns.

On the other hand, a PUSH approach, in which the ap-

plication proactively triggers a recommendation pro-

cess according to users’ preferences and when nec-

essary (e.g., when the user is close to a POI, when

lunch time is approaching, or when weather condi-

tions change) seems to be a more reasonable solution.

A ﬁrst step in this direction has been taken by Google

https://www.apptentive.com/blog/2017/06/22/how-many-

mobile-apps-are-actually-used/

https://www.simform.com/the-state-of-mobile-app-usage/

358

Castellanos, G., Cardinale, Y. and Roose, P.

Context-aware and Ontology-based Recommender System for e-Tourism.

DOI: 10.5220/0010552803580372

In Proceedings of the 16th International Conference on Software Technologies (ICSOFT 2021), pages 358-372

ISBN: 978-989-758-523-4

via Google Now

, which provides very basic infor-

mation/suggestions to users by detecting in their ge-

ographical environment what they need according to

their location and time. However, such applications

are simplistic and use only a limited amount of infor-

mation, and moreover, do not take into account the

user’s activity. A key element is the consideration of

the users’ context, such as their locations, weather (is

it raining?), or time (is it night?).

Recommender systems are now at the heart of

much research and offer real beneﬁts to users, orga-

nizations, and the business community (Lim et al.,

2019; Leskovec et al., 2020; Hamid et al., 2021).

Most tourism recommender systems are able to pre-

dict user’s preferences based on previous activities,

but research on Context-Aware Recommender Sys-

tems are still to be further developed (Laß et al., 2017;

Raza and Ding, 2019). Imagine a tourist in Paris, fan

of museums and parks, who wishes to spend the af-

ternoon in a good place, but it is a rainy day. The

experience should not be affected by the bad weather.

She/He would like to use an app that recommends a

nearby good place for spending this rainy afternoon,

preferably a museum instead of a park, hence spend-

ing no time searching for possible places to visit.

Moreover, if in the next day it is still raining (it rains a

lot in Paris!), the tourist would like to visit a different

museum or other indoor place.

Additionally, most recommender systems do not

offer a standard representation of the knowledge re-

lated to tourism (POI, users’ information, context

information, etc.) and do not consider to vary

recommendations– i.e., they become predictable for

same users, under the same situations. Thus, there

is still a real need for research and design of inno-

vative solutions in this ﬁeld. Nowadays, the use of

ontologies to represent such as knowledge is becom-

ing a powerful tool to offer such as innovative solu-

tions (Borr

as et al., 2014; Yochum et al., 2020).

To overcome these limitations, we propose RE-

CESO, a RECommender system architecture for E-

tourism with Serendipity and Ontology-based. RE-

CESO is a user-centric recommender system architec-

ture, that supports both PULL and PUSH approaches.

This framework is assisted by a semantic represen-

tation of the information managed, such as users’

preferences, their contexts, and POI, from which

an ontology-based spreading activation algorithm for

context-aware recommendations is designed. Ac-

cording to the users’ wishes, the serendipity of recom-

mendations may increase (e.g., surprise events, no re-

peated recommendations) (Kotkov et al., 2016), based

on an aging-like algorithm, which gives less priority

https://www.google.com/intl/fr/landing/now/

to recently recommended places to the user and in-

creasing priority of popular POI, even if they are not

in the user’s preferences.

We describe the architecture and algorithms of

RECESO and demonstrate its suitability and perfor-

mance, through a ﬁrst developed prototype of the ar-

chitecture, in different simulated scenarios, with dif-

ferent users’ proﬁles and preferences and different

context parameters. Results show how the spreading

activation and the proposed aging system give rele-

vant and varied recommendations, according to users’

preferences, while taking into account their context.

2 RELATED WORK

Depending on the technique used to make sug-

gestions, recommender systems are classiﬁed

into two categories: memory-based and model-

based (Bobadilla et al., 2013). Memory-based

recommender systems support their suggestions on

similarities among users and their shared items,

hence recommending similar items to the ones the

user likes (i.e., content-based recommendation) or

recommending items liked by users that are similar to

the user (i.e., collaborative ﬁltering). Model-based

recommender systems try to guess how much a

user will like an item that they did not consume

before, usually through statistical techniques and

machine-learning models. Recommender systems

that combine both approaches are called Hybrid

recommender systems.

We survey some relevant and recent studies in

the tourism domain, classifying them according to

these three categories and considering aspects related

to how they manage users’ preferences and interests,

the context awareness, the use of ontologies, and the

variability of recommendations. We compare them

in terms of these criteria and highlight the difference

with our proposal.

2.1 User Information

All recommender systems take into account some

information related to the user, such as identiﬁca-

tion aspects (sex, age, profession, etc.), interest and

preferences, or social relations. This information

can be obtained explicitly (e.g., questionnaires) (Laß

et al., 2017; Jannach and Zanker, 2020) or implic-

itly (e.g., by analyzing feedbacks in user’s social net-

works) (Lin et al., 2018).

Some proposals are supported on users’ inter-

action to obtain their data (i.e., following a PULL

paradigm). In (Rajaonarivo et al., 2019), authors

Context-aware and Ontology-based Recommender System for e-Tourism

359

model the users’ information considering their gen-

der, age category (e.g., kid, adult, or elderly), and

preferences, classiﬁed as thematic preferences (e.g.,

museum, theater) and historical preferences (e.g.,

12th century), that have to be provided by them

through a user interface. In (Santos et al., 2019), only

health, physical, and psychological conditions are re-

quired with forms ﬁlled by users. Users’ preferences

and feedback are directly assigned by the users in the

proposals presented in (Laß et al., 2017; Bahramian

et al., 2017). As in (Arigi et al., 2018), users’ inter-

est (i.e., not interested, less interested, and interested

enough) on tourism categories are asked to them.

SMARTMUSEUM (Ruotsalo et al., 2013), asks users

about the desired duration of a visit to a particular lo-

cation, the motivation for a visit, and ability to con-

sume the content offered by the system. The system

proposed in (Shen et al., 2016) requires photos up-

loaded by users, from which it extracts their travel

history; it also asks users to rank POI (their favorite

and non-favorite attractions).

Other recommender systems extract users’ infor-

mation, mainly their preferences and interests, from

available sources, without asking for explicit interac-

tion. The work presented in (Kesorn et al., 2017),

takes from Facebook basic information (e.g., name,

age) and check-in data (e.g., visited places) to iden-

tify users’ interests and preferences. To automat-

ically deduce users’ preferences from their social

networks, many techniques are used, such as opin-

ion mining (Logesh et al., 2018; Logesh and Sub-

ramaniyaswamy, 2019), analysis of implicit/explicit

feedback (Hidasi and Tikk, 2016), and analyzing geo-

tagged pictures (Sun et al., 2019). Curumim (Menk

et al., 2017) takes from users’ social networks their

travel history and level of education, and predicts their

degree of curiosity. Most of these works follow the

PULL approach.

Many other proposals combine both explicit and

implicit users’ information gathering, also under the

PULL paradigm. SPETA (Garc

ıa-Crespo et al.,

2009), supports its recommendation on the inter-

est and rating of tourism places that users explicitly

provide, and on preferences deduced by analyzing

their behavior on social networks. The system pre-

sented in (Alonso et al., 2012), takes into account

special needs and context-dependent preferences on

tourism sites, directly provided by users, as well as

explicit/implicit feedback on their social networks.

2.2 Context-awareness

In order to improve suggestions, the trend is to con-

sider context aspects that describe a specif situation

in a determined moment for a user, including trans-

portation media, weather, time, or even health condi-

tions. Many works use context-modeling approaches

that mainly consider means of transportation, travel

time, location, or weather (Bahramian et al., 2017;

Kesorn et al., 2017; Arigi et al., 2018; Logesh et al.,

2018; Rajaonarivo et al., 2019; Logesh and Subra-

maniyaswamy, 2019).

SMARTMUSEUM (Ruotsalo et al., 2013) only

considers users’ location as contextual information

captured by the built-in sensors of mobile devices,

such as GPS, accelerometers, and RFID readers. The

system proposed in (Shen et al., 2016) collects au-

tomatically the users’ current location (city, latitude,

and longitude) and current time, thus recommen-

dations are inﬂuenced by the geo-distance to POI.

SPETA (Garc

ıa-Crespo et al., 2009) also considers

users’ location extracted from users’ mobile devices,

but it takes into account current and the history of past

locations. It also gathers from mobile devices other

contextual information, such as weather forecast and

time. A similar work is proposed in (Laß et al., 2017),

that considers previously visited POI, time of the day,

day of the week, weather, temperature, and opening

hours to recommend a torist trip.

A General Factorization Framework for context-

aware recommender systems is proposed in (Hidasi

and Tikk, 2016), with the aim of constructing fac-

torization matrices (not matter which and how many

context factors are considered) for machine learning

techniques. In (Sun et al., 2019), it is used a combi-

nation of contextual information like weather, trans-

portation, or textual information, with geotagged pic-

tures taken by the user for building the user model.

The system proposed in (Alonso et al., 2012), is the

only one among the referenced works that considers

context-dependent preferences, as our proposal, re-

lated to weather, day, transport, and special needs. It

means that users’ preferences for POI, are expressed

with regards the context (e.g., indoor places when is

raining, art exhibitions at night).

2.3 Serendipity

A useful goal for a recommender system is to be

serendipitous, which means that the recommended

items must be relevant, novel, and unexpected. A rel-

evant item is an item that the user likes, consumes,

or is interested in; a novel item is one that users

have never seen or heard about in their life; an un-

expected item is signiﬁcantly different from the pro-

ﬁle of the user. Some quite good recent surveys em-

phasize the importance of such as feature for recom-

mender systems in general (Kotkov et al., 2016; Chen

ICSOFT 2021 - 16th International Conference on Software Technologies

360

et al., 2019) and in the tourism domain (Tintarev et al.,

2010; Menk et al., 2019).

It starts to be a trend to recommend POI. The

model considered in (Rajaonarivo et al., 2019), con-

siders previous activity of other users to recommend

tours of POI, thus reducing overspecialization and in-

creasing serendipity, however experiments with this

feature were not made. Authors of (Shen et al., 2016)

assert that the proposed system is able to recommend

fresh and surprise POI, based on collective intelli-

gence. From the level of curiosity predicted from

users’ social networks, Curumim (Menk et al., 2017)

adapts the degree of surprise and unexpectedness of a

recommended POI, tailored to users’ curiosity values.

2.4 Use of Ontologies

The huge amount of data that can be managed in

recommender systems, related to users’ information,

users’ preferences, context factors, POI, etc., de-

mands the use of more complex knowledge. Even

though, recommender system is an area that has been

the focus of many studies, thus reaching a very good

level of maturity, there is still a lack of standardiza-

tion to represent such information. In this sense, it

is evident the necessity of a well-deﬁned and stan-

dard model for representing the knowledge managed

by recommender systems. Semantic Web, in par-

ticularly the use of ontologies, seems to be a clear

solution, from which we can take its organizational

and relational capacity. In the context of tourism,

ontology-based recommender system is an emerging

trend (Borr

as et al., 2014; Yochum et al., 2020).

Some systems only count on ontologies to rep-

resent tourism POI (Garc

ıa-Crespo et al., 2009;

Bahramian et al., 2017; Arigi et al., 2018; Rajaonar-

ivo et al., 2019). Other works consider user proﬁle

ontologies, besides tourism ontologies, such in (Ruot-

salo et al., 2013). In (Alonso et al., 2012), an ontol-

ogy to represent user preferences and context factors

is proposed.

2.5 Discussion

Table 1 shows a comparative evaluation of the ref-

erenced works, in terms of category of the sys-

tem (model-based, hybrid, etc.), the access approach

(PULL or PUSH), type of user’s information gath-

ered, context factors considered, ontology used, how

they handle serendipity, and the information extrac-

tion model to obtain user and context data (explicit or

implicit). All cited recommender systems take into

account some information related to the user, such as

identiﬁcation aspects (sex, age, profession, etc.), in-

terest and preferences, or social relations, following

a PULL approach; thus, users ask for recommenda-

tions and have to control the system. Only few of

them apply techniques to analyze implicit informa-

tion, mainly from social media; hence, reducing the

intervention of users and being able of dynamically

updating the respective information, as we propose in

the RECESO architecture. Different types of context

data is considered for these recommender systems,

demonstrating the importance of such aspect. The

majority of them do not offer formal representation

of the knowledge managed with ontologies neither a

level of serendipity, as we do with RECESO.

The study presented by (Rajaonarivo et al., 2019)

is the only work that proposes a context-aware sys-

tem using ontologies and approaching serendipity, as

our proposed system RECESO, nonetheless it fol-

lows a PULL approach, the extraction of information

is based on explicit interaction of users, and authors

do not experiment the serendipitous feature. In con-

trast, RECESO supports the PUSH paradigm by im-

plicitly gathering users’ preferences and context and

automatically making recommendations (without ex-

pecting users asking), supports the PULL paradigm

by explicit user interactions, and we experimentally

evaluate the serendipity.

3 RECESO ARCHITECTURE

In order to overcome the limitations of existing rec-

ommender systems, we have identiﬁed the following

requirements for a new proposal:

• Support the PUSH paradigm, thus reducing the

user intervention to get recommendations.

• User-centric, thus returning more relevant items

to the users.

• Context-aware approach, for recommending more

suitable POIs.

• Ontology-based system, in order to offer an inter-

operable and ﬂexible system.

• Variability in recommendations to increase

serendipity and offer more diverse experiences.

• Combine the extraction of explicit and implicit in-

formation of users (e.g., from social media, their

smartphones), to manage up to date information.

RECESO can be accessed from a front-end that

implements either the PULL or PUSH paradigm or

both; it uses user’s preferences for recommending

relevant places; it uses contextual data gathered by

the front-end alongside user’s preferences for making

context-aware recommendations; a tourism ontology

is used for modeling the knowledge managed by the

Context-aware and Ontology-based Recommender System for e-Tourism

361

Table 1: Related work on recommender systems for e-tourism.

Reference Category/ Users’ Context Ontology Serendipity Information

PULL or PUSH information extraction

Garc

ıa-Crespo et al. Hybrid Preferences, Location Tourism Explicit,

(2009) PULL Social Implicit

Alonso et al. Hybrid Preferences, Weather, Transport, Tourism, Explicit,

(2012) PULL Social Day, Special Needs Context Implicit

Ruotsalo et al. Model based Duration, Moti- Location Tourism, Explicit

(2013) PULL vation, Ability User

Shen et al. Model based Preferences, Location, Time Explicit

(2016) PULL History

Hidasi and Tikk Model based Feedback, General Explicit,

(2016) PULL Social Implicit

Kesorn et al. Hybrid Social, Transportation, Implicit

(2017) PULL Check-in data Location, Weather

Bahramian et al. Hybrid Preferences Weather, Time, Day Tourism, Explicit

(2017) PULL Context

Laß et al. Memory based Preferences, Weather, Time, Day, Explicit

(2017) PULL Feedback Previously visited,

Opening hours

Menk et al. Hybrid Social, History, Curiosity Implicit

(2017) PULL Education

Arigi et al. Model based Preferences Transportation, Tourism Explicit

(2018) PULL Location, Weather

Logesh et al. Hybrid Social Transportation, Implicit

(2018) PULL Opinion mining Location, Weather

Rajaonarivo et al. Hybrid Basic, Transportation, Tourism Collaborative Explicit

(2019) PULL Preferences Location, Weather

Santos et al. Model based Health Explicit

(2019) PULL

Logesh and Hybrid Social, Transportation,

Subramaniyaswamy PULL Opinion Location, Weather Implicit

(2019) mining

Sun et al. Model based Social, Pictures Transportation, Explicit

(2019) PULL Textual Weather

RECESO Hybrid Preferences Weather, Time, Tourism, User, Aging Explicit,

PULL/PUSH Day, Location Context Implicit

recommender system (i.e., users’ preferences, context

factors, and POI); and it implements an aging system

for recommending the less frequent places more. The

general architecture of RECESO is mainly composed

by the follwing four modules, as shown in Figure 1:

• Data Gathering Module: User’s preferences are

received by the system, explicitly (i.e., direct in-

teraction) or implicitly (e.g., data mining, so-

cial networks analysis, user’s pictures analysis).

These preferences are related to categories of POI,

according to the POI ontology.

• User Interest Module: User’s preferences are

propagated from higher classes to lower classes

of the POI ontology.

• Context Module: The system receives informa-

tion about the context of the user, explicitly or im-

plicitly (retrieved from an API, mobile informa-

tion, etc.).

• Recommendation Module: The system recom-

mends a set of places to the user.

Figure 1: System architecture.

ICSOFT 2021 - 16th International Conference on Software Technologies

362

3.1 Data Gathering Module (User

Interface Client)

Initially, the system receives or calculates initial pref-

erences, as values between 0 and 1, for the higher

classes of the POI ontology. To obtain these values,

the user could explicitly set them by interacting with

the system or they could be implicitly determined us-

ing data mining, social network analysis, geo-data,

similar proﬁle users, friends preferences, etc. Dur-

ing the lifetime of the system, users’ preferences are

updated, according to their behaviors.

This module represents the client of the system

and can be implemented as a mobile apps, web apps,

or any other suitable front-end, which should be in

charge of supporting PULL or PUSH paradigms.

PULL approach demands users to explicitly interact

with the system to specify her/his data (e.g., personal

information, preferences, context) or to ask for a rec-

ommendation; while with the PUSH paradigm, the

front-end gathers information to send to the system,

which then returns the recommendation.

3.2 User Interest Model Module

Inspired on the work of Bahramian et al. (2017),

we introduce the concept of semantic network: a

tourism ontology extended with the user’s preferences

(see Figure 2) and context factor links (see Figure

4). We take advantage of the hierarchy of the se-

mantic network for propagating the preference of su-

perclasses to subclasses. Alongside preferences, each

node of the semantic network has a conﬁdence related

to the user, a value between 0 and 1, that deﬁnes how

sure is the system about the preference. When a pref-

erence is explicitly given by the user, its conﬁdence is

1, but when it is inferred from its ancestors or other

kind of analysis, its conﬁdence should be less than 1.

For each user, we compute the preference (pre f

)

and the conﬁdence (con f

) of each class c, accord-

ing to Eq. (1) and Eq. (2), respectively; where

ancestors(c) is the set of ancestors of the ontology

class c, pre f

is the user’s preference of the class c,

con f

is the conﬁdence about the user’s preference for

class c, and α is the decrease rate parameter, indicat-

ing how much the conﬁdence decreases at each level.

pre f

∑

p∈ancestors(c)

con f

pre f

∑

p∈ancestors(c)

con f

(1)

con f

∑

p∈ancestors(c)

con f

|ancestors(c)|

− α

(2)

These calculations are applied to each node

traversing from the higher classes, whose preferences

are obtained from the Data Gathering Module, to the

sinks, while decreasing conﬁdence values and there-

fore preference values. This process is called pref-

erence propagation. Figure 2 shows an example

of initial preferences gathered with values 0.85, 0.5,

and 0.7 for the ontology classes Cultural, Store, and

Sport, respectively. The initial conﬁdence for each

preference is 1, since they are gathered explicitly from

the user. Then, Figure 3 shows the preference prop-

agation with these values, applying Eq. (1) and Eq.

(2), with α = 0.1.

Figure 2: Initial preferences for Cultural, Store, and

Sport classes.

Figure 3: Preference propagation for Museum, Art

Gallery, Souvenirs, Stadium, and Golf classes, with

0.1 as decrease rate (α=0.1).

3.3 Context Module

We deﬁne the activation of a node of the semantic

network as the value that determines how feasible is

the user to visit a place that belongs to the category

of an ontology class, according to the current user’s

context. The user’s context is determined by the con-

text factors (e.g., time, day, weather, transportation,

mood), that could affect her/his decision to go to a

speciﬁc place. To make a recommendation, the sys-

tem has the context factors per user (contextFactors

set), obtained either explicit (provided by users) or

implicit (deduced by historical behaviors, data min-

Context-aware and Ontology-based Recommender System for e-Tourism

363

ing, etc.). For example, in Figure 4 the context of the

user is cloudless, night, and weekend.

Let’s deﬁne f

the fulﬁllment of a context factor x,

where f

= 1 if x fulﬁlls or f

= 0 otherwise, and r

c,x

as the relevance of x for the class c, which is an inte-

ger value in [0,2] that speciﬁes how much the context

factor can affect the user’s decision to go to a POI in c

(POI ontology class): 1 means indifference, 2 means

the fulﬁllment increases the wish to go to the POI,

0 means the fulﬁllment decreases the wish to go the

POI. Hence, we deﬁne act

as the activation for a class

c which is directly linked to the user’s context factors

(contextFactors), as shown in Eq. (3), and the activa-

tion for subclasses of c is deﬁned in Eq. (4).

act

∑

x∈contextFactors

c,x

(3)

act

∑

∈ancestors(c)

act

|ancestors(c)|

(4)

In the following, we show an example (Figure 4, Fig-

ure 5, and Figure 6), where explicit or implicit ob-

tained relevance are:

• r

Cultural,cloudless

= 1

• r

Cultural,night

= 0.5

• r

Cultural,weekend

= 1.8

• r

Store,cloudless

= 1.5

• r

Store,night

= 1.3

• r

Store,weekend

= 1

• r

Sport,cloudless

= 2

• r

Sport,night

= 0.2

• r

Sport,weekend

= 1.5

• f

cloudless

= f

night

= f

weekend

= 1

• f

= 0, x /∈ {cloudless,night,weekend}

First, the system deduces (from the user smart-

phone) the context of the user, which in this example

is cloudless night on the weekend (Figure 4). Then,

using Eq. (3), the activation values for Cultural,

Store, and Sport are computed (Figure 5). Finally,

using Eq. (4), the activation values for Museum, Art-

Gallery, Souvenirs, Stadium, and Golf are computed

(Figure 6). Also, this module detects the user’s loca-

tion, which is used for querying near POI.

3.4 Recommendation Module

Besides user’s preferences related to POI and context

factors, the system bases the recommendations on an

aging-like algorithm to ensure serendipity, on the dis-

tance of the POI to the user’s location and its trans-

portation medium, and on calculated scores of POI.

Figure 4: System receives cloudless, night, weekend

as user context.

Figure 5: Compute activation for Cultural, Store, and

Sport classes.

3.4.1 Aging System

Let’s deﬁne η

as the POI p’s aging, initialized as

= 1. Let’s deﬁne H as the aging rate. Each time

a POI p is recommended to the user, η

decreases by

H. When η

< 0.1, η

is reset to 1.

3.4.2 Great-circle Distance

Since the euclidean distance between two points on

Earth would cross through the surface, we should

use a more convenient measurement of distance: the

great-circle distance or orthodromic distance. It is

the shortest distance, along the surface of a sphere,

between two points on the surface of the sphere. It is

measured with circles on the sphere whose centers co-

incide with the center of the sphere. Those circles are

called great-circles. If we assume Earth is a perfect

sphere and hence use Great-Circle distance, we get

distances with errors no more than 0.5%, according

to (Ministry of Defense, London, 1997). The distance

between two points i and j on a sphere of radius r is

computed as shown in Eq. (5).

dist

i, j

= r·arccos(cos(lat

)·cos(lat

)·cos(lon

−lon

)+ sin(lat

)·sin(lat

))

(5)

ICSOFT 2021 - 16th International Conference on Software Technologies

364

Figure 6: Spread activation to Museum, Art Gallery,

Souvenirs, Stadium, and Golf classes.

3.4.3 Score

The system calculates the score of each instance p of a

class c in the ontology, in terms of user’s preferences

(Eq. (1)), activation level (Eq. (3) or Eq. (4)), the

aging value (η

), and the great-circle distance. Eq.

(6) shows how the score is calculated; where w

are

conﬁguration parameters of the system that represent

the weigh of each term (

∑

= 1), pre f

is the user’s

preference on the class c, act

is the activation level of

class c, η

is the aging factor of the POI p, dist

u,p

the great-circle distance between user’s location and

POI p, and maxdist is the maximum great-circle dis-

tance that a p should be from user u (this distance de-

pends on the user transportation that can be detected

from the user’s smartphone sensors, for example).

score

= w

· pre f

+ w

· act

+ w

· η

− w

dist

u,p

maxdist

(6)

The system uses this score for returning the list of

”top places”, whose length could be conﬁgured.

4 STUDY CASE: EXPERIMENTS

AND EVALUATION

To evaluate our system, we implemented a proto-

type

, with Java, a Fuseki server as the triplestore

to manage the semantic repository with the ontology,

Apache Jena for querying the ontology and connect-

ing to the triplestore, and a MySQL database for stor-

ing additional information. To have initial data, we re-

quested real users through questionnaires, whose in-

formation was used to create simulated users and sce-

narios for the tests. With this version of RECESO, we

https://github.com/gustavoaca1997/datatourisme-

recommender

evaluate the User Interest Model Module, the Recom-

mendation Module, and the Activation Propagation

algorithm from the Context Module, while the Data

Gathering Module and the context gathering (Loca-

tion and Context Factors) are simulated.

4.1 Ontology Classes

We consider a light version of DATATourisme

ontology

, an open data ontology to manage tourism

in France, consisting of the Point of interest class, its

Place subclass, from which we take: Cultural Site,

Food Establishment, Leisure Place, Natural Heritage,

Sports, and Store. We slightly modiﬁed the DATA-

Tourisme ontology to specialize the class Sports and

Leisure Place into separated Sports class and Leisure

Place class, hence making less ambiguous what kind

of places should belong to each category. Figure

7 shows the subset of classes of DATATourisme

ontology, considered in the experiments.

The high level classes that are directly linked to

the context factors values are:

• Museum

• Interpretation

Center

• Library

• Park and Garden

• Archaeological

Site

• Religious Site

• Remarkable Build-

ing

• City Heritage

• Defense Site

• Remembrance Site

• Technical Her-

itage

• Food Establish-

ment

• Natural Heritage

• Sports

• Leisure Place

• Store

Figure 7: Subset of the modiﬁed version of the ontology

DATATourisme.

http://info.datatourisme.gouv.fr/ontology/core/2.0/

Context-aware and Ontology-based Recommender System for e-Tourism

365

4.2 From Real Users to Simulated

Scenarios

In order to test several scenarios with different user’s

preferences, proﬁles, and behaviors, we created

synthetic scenarios, however based on data gathered

from real people. We asked people to ﬁll a google

form with two parts:

(1) To Determine the Relevance for each context

factor value. The list of asked context factors included

weather (with possible values rainy, cloudless,

snowy), time (with possible values early morning,

morning, afternoon, night), and day (with pos-

sible values weekday, weekend). The list of asked

types of POI was the high level classes listed in the

previous section. The question was ”how much the

context factor x inﬂuences your decision to go to a

’place/event’ of the type c (these types are not ex-

clusive)”. People should answer an integer value be-

tween 0 and 2, where:

• 0 means that if the context factor is met, the user

would not go to the place/event.

• 1 means the user does not care if the context factor

is met or not.

• 2 means that if the context factor is met, the user

would go to the place/event.

Then, for each pair of ontology class, c, and

context factor, x, (for example Museum with rainy),

the average relevance is computed and stored as the

relevance for the experiments.

(2) To Determine Users’ Preferences for tourism

categories. In this part of the questionnaires, peo-

ple provided their preferences of the high level

ontology classes (representing POI), and also the

genre, country, profession, age, and (optional) social

networks to have more information for further work.

We got 102 answers from people of different ages,

professions, and countries. With this information we

generate synthetic tourists, with different behaviors

for different context scenarios.

Simulated Scenarios. To evaluate the system, we test

it with simulated scenarios (simulated contexts and

simulated users). The data gathered with question-

naires were grouped into four cluster, with the Elbow

Method (increasing the number of clusters does not

make improvements worth the cost using our small

dataset). The centroids of the four clusters repre-

sent our four simulated tourists, T

. Table 2

shows the preferences for each tourist to each tourism

category, according to our light version of DATA-

Tourise ontology. Figure 8 shows for each tourist how

the preferences are distributed, where we can see that

is the one with a more varied set of preferences,

including the lowest values between all preferences.

Table 2: Simulated tourists’ preferences.

Class T

Museum 0.7235 0.6476 0.78 0.9421

Interpretation Center 0.6824 0.6476 0.52 0.8737

Library 0.5588 0.46667 0.72 0.7684

Park and Garden 0.865 0.8714 0.7 0.9316

Archaeological Site 0.6882 0.8619 0.82 0.826

Religious Site 0.45882 0.7 0.36 0.7316

Remarkable Building 0.6529 0.8524 0.4 0.7895

City Heritage 0.7412 0.919 0.58 0.8632

Defence Site 0.5412 0.8529 0.68 0.7421

Remembrance Site 0.4941 0.781 0.8 0.7211

Technical Heritage 0.4059 0.7905 0.56 0.5842

Food Establishment 0.9059 0.8905 0.7 0.6632

Natural Heritage 0.9059 0.9143 0.76 0.9211

Sports 0.9353 0.8429 0.22 0.6316

Leisure Place 0.9353 0.8429 0.22 0.6316

Store 0.7647 0.8286 0.28 0.5684

Figure 8: Distributions of initial preferences of T

, T

and T

For our experiments, all tourists have the same

relevance values, as we show in Table 3. For each

tourist T

, we simulate visits to Paris, Niza, and Lyon.

For each combination of context factors (24 combina-

tions), each T

goes twice to each site; thus 48 visits

(see Figure 9).

4.3 Metrics

Let R be a recommendation set for a speciﬁc user on a

speciﬁc location and a speciﬁc context (combination

of context factors), we deﬁne some metrics to evaluate

how ”good” is a recommendation R, as follows:

• Average preference of R (Eq. (7)), which denotes

how near to the user’s preferences are the recom-

mended POIs.

pre f

∑

p∈R

pre f

|R|

(7)

ICSOFT 2021 - 16th International Conference on Software Technologies

366

Table 3: Simulated relevance of context factors for T

Class rainy cloudless snowy workday weekend morning afternoon night early morning

Museum 1.0 1.394 1.029 1.082 1.488 1.065 1.518 0.788 0.194

Interpretation Centre 0.871 1.312 0.829 1.029 1.424 1.029 1.382 0.771 0.135

Libreary 1.271 1.388 1.188 1.312 1.159 1.324 1.453 0.729 0.271

Park and garden 0.335 1.847 0.894 1.365 1.665 1.388 1.629 0.924 0.494

Archeological Site 0.476 1.488 0.671 0.782 1.471 1.235 1.424 0.682 0.335

Religious Site 0.8 1.018 0.841 0.641 1.106 1.018 0.971 0.529 0.1

Remarkable building 0.912 1.576 0.994 1.165 1.506 1.241 1.453 1.0056 0.324

City heritage 0.665 1.629 0.947 1.088 1.618 1.329 1.565 1.271 0.5

Defense Site 0.712 1.412 0.806 0.835 1.376 1.253 1.388 0.824 0.265

Remembrance Site 0.588 1.318 0.812 0.935 1.306 1.129 1.294 0.6 0.324

Technical Heritage 0.582 1.371 0.682 0.841 1.3 1.265 1.3 0.741 0.3

Food Stablishment 1.488 1.624 1.494 1.635 1.759 1.353 1.671 1.676 0.688

Natural Heritage 0.494 1.776 0.835 1.088 1.659 1.535 1.524 0.953 0.465

Leisure place 1.088 1.541 1.006 1.118 1.588 1.024 1.506 1.418 0.588

Sports 0.1 2.0 0.75 1.118 1.588 1.024 1.506 0.8 0.588

Store 1.312 1.535 1.241 1.388 1.435 1.0588 1.547 1.1647 0.3

Figure 9: Simulated visits.

• Average activation of R (Eq. (8)), which denotes

how relevant to the user are the current context

factors to the recommended POIs.

act

∑

p∈R

act

|R|

(8)

• Average aging of R (Eq. (9)), which denotes how

”aged” to the user are the recommended POIs.

∑

p∈R

|R|

(9)

• Average distance of R (Eq. (10)), which denotes

how far are the recommended POIs from the user.

dist

∑

p∈R

dist

|R|

(10)

• Average novelty of R (Eq. (12)), which denotes

how novel are the recommended POIs to the user.

In Eq. (12), nov

is calculated by Eq. (11), which

is based on the work of Kotkov et al. (Kotkov

et al., 2016), where p is a recommended place, rec

is the set of already recommended places to the

user, c

is the ontology class to which p belongs,

and dist computes the distance of two classes in

the ontology (using Breadth First Search).

nov

q∈rec

min (dist(c

))

(11)

nov

∑

p∈R

nov

|R|

(12)

4.4 Experiments

To calculate the score for each POI p, we need some

user-deﬁned parameters, that represent maximum dis-

tance (maxdist) from the user to the POI and the

weighs (w

) for each term in Eq. (6) (see Section

3.4.3). These user-deﬁned parameters are speciﬁed

just once, at the installation/conﬁguration of the sys-

tem. We designed two scenarios with different values

of these conﬁguration parameters, as we explain in

the following.

4.4.1 First Conﬁguration

We consider maxdist = 8Km, w

= 0.3611 for prefer-

ence, w

= 0.3611 for activation, w

= 0.25 for aging,

and w

= 0.0278 for distance from user, resulting on

Eq. (13). In this conﬁguration, we give more prior-

ity to preference and activation and very little priority

to distance from user. With these parameters we start

simulating each user and their visits.

score

= 0.3611 · pre f

+ 0.3611 · act

+ 0.25 · η

− 0.0278 ·

dist

u,p

(13)

Context-aware and Ontology-based Recommender System for e-Tourism

367

For space limitations, we are showing only some

sets of recommended items like Table 4, where each

is a recommended place and score

≥ score

for

every i < j. Table 4 corresponds to T

’s ﬁrst visit

to Niza, on a rainy workday in the morning, where

is Train des Merveilles (from class TouristTrain),

is Cin

ema de Plein Air (from class Cinema), p

Casino de Beaulieu (from class Casino), p

is Cin

ema

de Beaulieu (from class Cinema), and p

is Lyon-style

Petanque Fields (from class BoulesPitch). Note that

since all places are instances of LeisurePlace class,

which T

loves, they have same preference and acti-

vation; also, because this is the ﬁrst recommendation,

all aging values are 1.0. Therefore, the distance is the

tiebreaker. Despite of a rainy morning context, the

system recommends a Tourist Train, just because it is

a Leisure Place.

Table 4: First visit of T

to Niza with ﬁrst conﬁguration and

context = (rainy, morning, workday).

Field p

pre f

0.9353 0.9353 0.9353 0.9353 0.9353

act

4.4 4.4 4.4 4.4 4.4

1.0 1.0 1.0 1.0 1.0

dist

u,p

1.53 3.99 5.52 5.53 5.58

score

2.1713 2.1628 2.15745 2.15742 2.1573

At the second visit to Lyon (see Table 5), on a

rainy workday in the morning, only p

is not a Leisure

Place but an Interpretation Center, a kind of place T

does not like as much as leisure places. However, the

activation of InterpretationCenter is higher enough to

make score

greater than score

. Despite of p

be-

ing very far away from T

’s location, the little magni-

tude of w

makes it little important to the ﬁnal score.

Something odd on the recommended set is that p

and

are the same place, Le Sucre, but reported as be-

longing to two different ontology classes: NightClub

and Theatre.

Table 5: Second visit of T

to Lyon with ﬁrst conﬁguration

and context = (rainy, morning, workday).

Field p

pre f

0.676 0.9353 0.9353 0.9353 0.9353

act

4.818 4.51 4.51 4.51 4.51

1.0 1.0 1.0 1.0 1.0

dist

u,p

7.1 4.45 4.54 4.54 4.61

score

2.2092 2.2015 2.2012 2.2012 2.2010

At the fourth of visit T

to Niza, on a rainy work-

day in the early morning, the system recommends

what is shown in Table 6, where p

is Cin

ema de

Plein Air (from class Cinema), p

is Ferronnerie d’art

JC Rodriguez (from class CraftsmanShop), p

is Ate-

lier Hesperida (from class CraftsmanShop), p

is Hor-

logerie Folt

ete (from class CraftsmanShop), and p

Galerie Bizet (from class CraftsmanShop). Since T

loves leisure places more than stores, the predicted

preference for Cin

ema de Plein Air is greater than the

other ones on this visit. However, we can see that

lowers down score

to a value not so different

from score

. The recommendations to T

returned

after this one do not have Cin

ema de Plein Air in the

recommended POIs until the context snowy workday

afternoon is reached in the loop, because of the aging

of Cin

ema de Plein Air and its low activation.

Table 6: Fourth visit of T

to Niza with ﬁrst conﬁguration

and context = (rainy, early morning, workday).

Field p

pre f

0.9353 0.7647 0.7647 0.7647 0.7647

act

4.2 4.1 4.1 4.1 4.1

0.6 1.0 1.0 1.0 1.0

dist

u,p

3.99 5.21 5.25 5.72 5.737

score

1.9906 1.9886 1.9884 1.98684 1.98678

While the system recommends same places to T

and T

for their ﬁrst visits to Niza, Lyon, and Paris, at

rainy workdays at afternoon, the sets start to diverge

after the second visits. The system recommends to T

a set of Leisure Places and one Interpretation Center

for the second visit to Lyon, but recommends a set

with two Archeological Sites, one Interpretation Cen-

ter, and two Remarkable Buildings to T

, in that order

(see Table 7). The more diverse initial preferences of

and the aging system are responsible for this.

Table 7: Second visit of T

to Lyon with ﬁrst conﬁguration

and context = (rainy, afternoon, workday).

Field p

pre f

0.852 0.852 0.659 0.843 0.843

act

4.659 4.659 4.818 4.6 4.6

1.0 1.0 1.0 1.0 1.0

dist

u,p

3.967 5.267 7.0969 3.967 5.267

score

2.226 2.222 2.203 2.202 2.197

After many visits, again at a rainy workday at af-

ternoon, the system recommends a completely differ-

ent set of places to T

when visiting Lyon. It recom-

mends a Theater, an Art Gallery, a Craftsman Shop,

and two other Art Galleries, in that order (see Ta-

ble 8). The system aging is responsible for this variety

of recommendations.

With a low value of w

, which represents aging’s

weight, there are cases where the recommendations

are not very diverse, as is the case of the two visits

of T

to Paris on cloudless workdays at early morn-

ing, shown in Table 9 and Table 10, where the sec-

ond one represents the recommendations made after

many visits. Both visits involve Garnier Opera (from

ICSOFT 2021 - 16th International Conference on Software Technologies

368

Table 8: A visit of T

to Lyon with ﬁrst conﬁguration after

many visits and context = ( rainy, afternoon, workday).

Field p

pre f

0.843 0.829 0.829 0.829 0.829

act

4.512 4.371 4.371 4.371 4.371

0.8 1.0 1.0 1.0 1.0

dist

u,p

7.157 5.425 5.438 5.442 5.462

score

2.10876 2.10864 2.10859 2.10858 2.10851

class Palace), Conciergerie (from class Palace), and

Le Manoir de Paris (from class InterpretationCentre),

as p

, p

, and p

for the ﬁrst visit and p

, p

, and p

for the second visit, respectively.

Table 9: First visit of T

to Paris with ﬁrst conﬁguration and

context = (cloudless, early morning, workday).

Field p

pre f

0.829 0.843 0.843 0.659 0.659

act

4.171 4.4 4.4 4.571 4.571

0.8 0.4 0.4 0.4 0.4

dist

u,p

7.51 4.25 5.06 5.9 6.3

score

1.97916 1.97869 1.97590 1.96803 1.96665

Table 10: Second visit of T

to Paris with ﬁrst conﬁguration

and context = (cloudless, early morning, workday).

Field p

pre f

0.843 0.843 0.659 0.659 0.659

act

4.40 4.40 4.57 4.57 4.57

0.4 0.4 0.4 0.4 0.4

dist

u,p

4.25 5.06 3.79 4.55 5.9

score

1.97869 1.97590 1.97535 1.97273 1.96803

Table 11 shows the average of the recommenda-

tion metrics for each tourist using the ﬁrst conﬁgu-

ration. Averages of pre f

of T

’s and T

’s visits are

the highest ones, being greater than 0.8. Next is T

’s,

being greater than 0.7, which is still a good average

preference. T

’s average pre f

is lower, but we can

blame the amount of low initial preferences of T

, as

we can see on Figure 8.

Table 11: Averages of recommendations metrics with ﬁrst

conﬁguration.

avg T

avg(pre f

) 0.8555 0.8213 0.4183 0.7269

avg(act

) 4.286 4.3092 4.314 4.3319

avg(η

) 0.7058 0.7111 0.6906 0.6914

avg(nov

u,R

) 0.0923 0.1231 0.1329 0.1147

avg(dist

u,R

) 4.876 4.8993 4.9101 4.9861

The average act

of each tourist is greater than

4.0. Theoretically, the maximum possible value of

act

is 6.0, but since it depends of the activation value

act

of the recommended ontology classes (see Eq.

(8)), which depends of the relevances r

c,x

of the user

(see Eq. (4)), that do not reach their maximum possi-

ble value 2.0 in the survey (see Section 4.2), it is im-

possible for these experiments to reach the maximum

act

. Therefore, we can say that an average act

4.0 is good enough.

Average η

is very near 0.7, hence the amount of

”young” recommended places is high. Theoretically,

the maximum possible value is 1.0, but that would be

possible if each recommendation involves POI never

recommended before.

Despite of recommending places with distance

from user near 8.0 km, the average dist

u,R

for each

tourist is acceptable: greater than 4.0 km and less than

5.0 km. The novelties measured as we proposed on

Section 4.3 have bad performance on four tourists.

4.4.2 Second Conﬁguration

Now we consider w

= 0.25 for preference, w

= 0.25

for activation, w

= 0.472 for aging, and w

= 0.0278

for distance from user, resulting on equation 14 and

giving now more priority to aging.

score

= 0.25 · pre f

+ 0.25 · act

+ 0.472 · η

− 0.0278 ·

dist

u,p

(14)

For the ﬁrst visit of T

to Niza in a cloudless work-

day at afternoon, the η

is 0.84 (see Table 13), while

it is 0.80 using the ﬁrst conﬁguration (see Table 12).

The pre f

improves, from 0.765 to 0.799, and the

act

goes from 4.44 to 4.33.

Table 12: First visit of T

to Niza with ﬁrst conﬁguration

and context = (cloudless, afternoon, workday).

Field p

pre f

0.76 0.76 0.76 0.76 0.76

act

4.44 4.44 4.44 4.44 4.44

0.8 0.8 0.8 0.8 0.8

dist

u,p

5.25 5.72 5.74 5.78 5.79

score

2.06166 2.06004 2.05998 2.05983 2.05981

Table 13: First visit of T

to Niza with second conﬁguration

and context = (cloudless, afternoon, workday).

Field p

pre f

0.76 0.76 0.935 0.76 0.76

act

4.44 4.44 3.89 4.44 4.44

0.8 0.8 1.0 0.8 0.8

dist

u,p

5.20 5.25 5.30 5.72 5.73

score

1.6612 1.6610 1.6597 1.6594 1.6593

For the second visit of T

to Paris on a snowy

weekend at early morning, the η

is 0.6 using second

conﬁguration (see Table 15), while it is 0.32 using

ﬁrst conﬁguration (see Table 14). This makes sense

since we give more weight to aging in the second con-

ﬁguration. The other metrics of R are very similar for

both visits.

Context-aware and Ontology-based Recommender System for e-Tourism

369

Table 14: Second visit of T

to Paris with ﬁrst conﬁguration

and context = (snowy, early morning, weekend).

Field p

pre f

0.866 0.866 0.790 0.866 0.866

act

4.288 4.288 4.265 4.288 4.288

0.4 0.4 0.4 0.2 0.2

dist

u,p

5.899 6.297 4.253 3.791 4.546

score

1.941 1.939 1.911 1.898 1.895

Table 15: Second visit of T

to Paris with second conﬁgura-

tion and context = (snowy, early morning, weekend).

Field p

pre f

0.866 0.866 0.738 0.866 0.866

act

4.288 4.288 3.81 4.288 4.288

1.0 1.0 0.6 0.2 0.2

dist

u,p

3.791 4.546 5.333 5.899 6.297

score

1.748 1.745 1.402 1.363 1.361

Table 16 shows the average of the recommenda-

tion metrics for each tourist using the second conﬁgu-

ration. Again, T

’s and T

’s avg(pre f

) are the high-

est ones, followed by T

’s, while T

’s is the lowest.

We can notice that only T

’s and T

’s avg(pre f

) in-

crease a little from the ones with ﬁrst conﬁguration,

but the other two decrease. Again, all avg(act

) are

above 4.0, but each one decreases from the ﬁrst con-

ﬁguration version. As expected, all avg(η

) increase

because the second conﬁguration gives more weight

to aging.

Table 16: Averages of recommendations metrics with sec-

ond conﬁguration.

avg T

avg(pre f

) 0.8658 0.8272 0.3912 0.6987

avg(act

) 4.0738 4.0818 4.0031 4.0415

avg(η

) 0.7592 0.7619 0.7850 0.7739

avg(nov

u,R

) 0.172 0.1860 0.1986 0.1580

avg(dist

u,R

) 4.842 5.0215 4.8999 4.9615

4.4.3 Comparison with Related Work

We implemented the system proposed in (Bahramian

et al., 2017) and executed it using the same score

function and same ﬁrst conﬁguration, we obtain the

average metrics shown in the Table 17. The average

pre f

for each T

are very similar; likewise for the

average activation values. The average nov

of RE-

CESO are higher, showing how our proposal improve

serendipity.

When we use the second conﬁguration, we ob-

tain the average metrics shown in the Table 18. We

can see that for T

and T

, the average preferences

returned by RECESO are better, but for T

and T

the average preferences returned by the system pro-

posed in (Bahramian et al., 2017) are better. The av-

erage activation values are worse for RECESO, but as

expected, the average nov

is signiﬁcantly higher for

RECESO, alongside the average η

4.4.4 Discussion

For measuring novelty we use Eq. (11) based on

Kotkov et al. (2016), since when novelty increases,

so does serendipity. That equation was proposed with

the hypotheses that if a user already knows places

from class c, then it is probable the user already

knows other places from class c. With that hypothe-

ses, our system gives low novelty at ﬁrst glance:

when a place belonging to the ontology class c is rec-

ommended, next recommendations with places from

c will have lower novelty. But when we consider

the comparison with the work of Bahramian et al.

(2017), the average novelty of RECESO is signiﬁ-

cantly higher, implying a better serendipity.

Experiments with the second conﬁguration are fo-

cused on increasing average aging by giving a higher

value to w

on Eq. (6), but the increment of aging is

paid with a decrease on activation. Nonetheless, two

simulated tourists have an increment for their pref-

erences, because on ﬁrst conﬁguration the system is

trapped between very activated POI but not very rel-

evant, but the heavier aging of the second conﬁgura-

tion increases variability, hence lets the system rec-

ommend other POI despite of lowering the activation.

Nevertheless, our system lets users choose each w

of Eq. (6), that way choosing if they want recommen-

dations mainly relevant than contextually activated

or varied, or even recommendations with the nearest

places ignoring any other attribute, or any other pos-

sible conﬁguration.

Comparing the proposed architecture of RECESO

with state-of-the-art studies (see Section 2), we over-

come some of their limitations: (i) combining PULL

and PUSH paradigms reduces users intervention and

control over the system and allows making recom-

mendations even if users do not ask the system (e.g.,

when users are near POI, when the weather change);

(ii) analyzing implicit information from users’ social

media or smartphones makes possible to dynamically

update users and context information; thus, man-

age most current data; (iii) Managing semantic in-

formation with ontologies to represent touristic POI,

user preferences, and context information, ensures a

ﬂexible and formal representation of the knowledge,

which in turn enables the implementation of more in-

telligent and interoperable recommender systems as

an alternative of machine learning based approaches;

and (iv) as shown by results, the aging-like algorithm

improves serendipity, compared with a state-of-the-

art work.

This ﬁrst prototype of RECESO demonstrates the

ICSOFT 2021 - 16th International Conference on Software Technologies

370

Table 17: Averages of recommendations metrics with ﬁrst conﬁguration using RECESO and Bahramian et al. system.

avg T

RECESO Bahramian RECESO Bahramian RECESO Bahramian RECESO Bahramian

avg(pre f

) 0.8555 0.832 0.8213 0.814 0.4183 0.4290 0.7269 0.74167

avg(act

) 4.286 4.347 4.3092 4.368 4.314 4.3558 4.3319 4.366

avg(η

) 0.7058 0.623 0.7111 0.620 0.6906 0.6247 0.6914 0.6206

avg(nov

) 0.0923 0.0699 0.1231 0.0853 0.1329 0.0573 0.1147 0.0741

avg(dist

) 4.876 4.359 4.8993 4.520 4.9101 4.7487 4.9861 4.761

Table 18: Averages of recommendations metrics with second conﬁguration using Bahramian et al. system.

avg T

RECESO Bahramian RECESO Bahramian RECESO Bahramian RECESO Bahramian

avg(pre f

) 0.8658 0.7416 0.8272 0.8144 0.3912 0.4292 0.6987 0.7416

avg(act

) 4.0738 4.366 4.0818 4.3678 4.0031 4.3556 4.0415 4.3660

avg(η

) 0.7592 0.6194 0.7619 0.6211 0.7850 0.6247 0.7739 0.61944

avg(nov

) 0.172 0.07413 0.1860 0.0853 0.1986 0.06154 0.1580 0.07413

avg(dist

) 4.842 4.7292 5.0215 4.5136 4.8999 4.7441 4.9615 4.7292

feasibility of a recommender system with all these ad-

vantages over existing studies. This experience also

gives the opportunity of extracting its current limita-

tions and some lessons learnt, as analyzing the score

and aging systems to improve serendipity and nov-

elty.

5 CONCLUSIONS AND FUTURE

WORK

In this paper we present RECESO, a context-aware

and ontology-based recommender system that is able

to recommend varied Points of Interest (POI). We use

a spreading activation algorithm to model users and

their context supported on an ontology that represents

POI and users’ preferences. We use a proposed ag-

ing system which gives more priority to POIs that are

not frequently recommended. To evaluate the perfor-

mance of RECESO, we perform some experiments,

by simulating visits of four simulated tourists, based

on real preferences obtained through a survey. Re-

sults show that it is able to recommend different sets

of places to the same user at the same context at differ-

ent moments, still taking into account the user’s pref-

erences and improving novelty level. We are develop-

ing a more complete version of RECESO, including

all its modules. We also plan to integrate it into mo-

bile or web apps able to gather user information and

contextual information, including location. We expect

to perform more validations with real scenarios.

ACKNOWLEDGEMENT

This research was partially funded by FONDO NA-

CIONAL DE DESARROLLO CIENT

IFICO, TEC-

NOL

OGICO Y DE INNOVACI

ON TECNOL

OGICA

- FONDECYT as executing entity of CONCYTECun-

der grant agreement no.01-2019-FONDECYT-BM-

INC.INV in the project RUTAS: Robots for Urban

Tourism Centers, Autonomous and Semantic-based.

REFERENCES

Alghamdi, H., Zhu, S., and El Saddik, A. (2016). E-

tourism: mobile dynamic trip planner. In Intl. Sym-

posium on Multimedia, pages 185–188. IEEE.

Alonso, K., Zorrilla, M., I

nan, H., Palau, M., Confalonieri,

R., V

azquez-Salceda, J., Calle, J., and Castro, E.

(2012). Ontology-based tourism for all recommender

and information retrieval system for interactive com-

munity displays. In Intl. Conf. on Information Science

and Digital Content Technology, volume 3, pages

650–655. IEEE.

Arigi, L. R. H., Baizal, Z. A., and Herdiani, A. (2018).

Context-aware recommender system based on ontol-

ogy for recommending tourist destinations at ban-

dung. In Journal of Physics: Conference Series, vol-

ume 971. IOP Publishing.

Artemenko, O., Kunanets, O., and Pasichnyk, V. (2017).

E-tourism recommender systems: a survey and devel-

opment perspectives. Econtechmod.

Bahramian, Z., Abbaspour, R. A., and Claramunt, C.

(2017). A context-aware tourism recommender sys-

tem based on a spreading activation method. ISPRS

- International Archives of the Photogrammetry, Re-

mote Sensing and Spatial Information Sciences, XLII-

4/W4:333–339.

Bobadilla, J., Ortega, F., Hernando, A., and Guti

errez, A.

Context-aware and Ontology-based Recommender System for e-Tourism

371

(2013). Recommender systems survey. Knowledge-

based systems, 46:109–132.

Borr

as, J., Moreno, A., and Valls, A. (2014). Intelligent

tourism recommender systems: A survey. Expert Sys-

tems with Applications, 41(16):7370–7389.

Chen, L., Yang, Y., Wang, N., Yang, K., and Yuan, Q.

(2019). How serendipity improves user satisfaction

with recommendations? a large-scale user evaluation.

In The World Wide Web Conference, pages 240–250.

Garc

ıa-Crespo, A., Chamizo, J., Rivera, I., Mencke,

M., Colomo-Palacios, R., and G

omez-Berb

ıs, J. M.

(2009). Speta: Social pervasive e-tourism advisor.

Telematics and informatics, 26(3):306–315.

Hamid, R. A., Albahri, A., Alwan, J. K., Al-qaysi, Z., Al-

bahri, O., Zaidan, A., Alnoor, A., Alamoodi, A., and

Zaidan, B. (2021). How smart is e-tourism? a sys-

tematic review of smart tourism recommendation sys-

tem applying data management. Computer Science

Review, 39:100337.

Hidasi, B. and Tikk, D. (2016). General factorization frame-

work for context-aware recommendations. Data Min-

ing and Knowledge Discovery, 30(2):342–371.

Jannach, D. and Zanker, M. (2020). Interactive and context-

aware systems in tourism. Handbook of e-Tourism,

pages 1–22.

Kayumovich, K. O. (2020). Prospects of digital tourism

development. Economics, (1 (44)).

Kazandzhieva, V. and Santana, H. (2019). E-tourism:

Deﬁnition, development and conceptual framework.

Turizam: medunarodni znanstveno-strucni

casopis,

67(4):332–350.

Kesorn, K., Juraphanthong, W., and Salaiwarakul, A.

(2017). Personalized attraction recommendation sys-

tem for tourists through check-in data. IEEE Access,

5:26703–26721.

Kotkov, D., Wang, S., and Veijalainen, J. (2016). A survey

of serendipity in recommender systems. Knowledge-

Based Systems, 111:180–192.

Laß, C., Herzog, D., and W

orndl, W. (2017). Context-

aware tourist trip recommendations. In Workshop on

Recommenders in Tourism co-located with 11th ACM

Conference on Recommender Systems, pages 18–25,

Como, Italy.

Leskovec, J., Rajaraman, A., and Ullman, J. D. (2020). Min-

ing of massive data sets. Chapter 9: Recommendation

Systems. Cambridge university press.

Lim, K. H., Chan, J., Karunasekera, S., and Leckie, C.

(2019). Tour recommendation and trip planning us-

ing location-based social media: a survey. Knowledge

and Information Systems, pages 1–29.

Lin, C.-Y., Wang, L.-C., and Tsai, K.-H. (2018). Hybrid

real-time matrix factorization for implicit feedback

recommendation systems. IEEE Access, 6:21369–

21380.

Logesh, R. and Subramaniyaswamy, V. (2019). Exploring

hybrid recommender systems for personalized travel

applications. In Cognitive informatics and soft com-

puting, pages 535–544. Springer.

Logesh, R., Subramaniyaswamy, V., and Vijayakumar, V.

(2018). A personalised travel recommender sys-

tem utilising social network proﬁle and accurate gps

data. Electronic Government, an International Jour-

nal, 14(1):90–113.

Menk, A., Sebastia, L., and Ferreira, R. (2017). Cu-

rumim: A serendipitous recommender system for

tourism based on human curiosity. In Intl. Conf.

on Tools with Artiﬁcial Intelligence, pages 788–795.

IEEE.

Menk, A., Sebastia, L., and Ferreira, R. (2019). Recommen-

dation systems for tourism based on social networks:

A survey. arXiv preprint arXiv:1903.12099.

Ministry of Defense, London (1997). Admiralty Manual

of Navigation: BR 45(1). Number v. 1 in BR Series.

Stationery Ofﬁce.

Rajaonarivo, L., Fonteles, A., Sallaberry, C., Bessagnet, M.-

N., Roose, P., Etcheverry, P., Marquesuza

a, C., La-

cayrelle, A. L. P., Cay

e, C., and Coudert, Q. (2019).

Recommendation of heterogeneous cultural heritage

objects for the promotion of tourism. ISPRS Interna-

tional Journal of Geo-Information, 8(5):230.

Raza, S. and Ding, C. (2019). Progress in context-aware

recommender systems—an overview. Computer Sci-

ence Review, 31:84–97.

Ruotsalo, T., Haav, K., Stoyanov, A., Roche, S., Fani, E.,

Deliai, R., M

akel

a, E., Kauppinen, T., and Hyv

onen,

E. (2013). Smartmuseum: A mobile recommender

system for the web of data. Journal of Web Seman-

tics, 20:50–67.

Santos, F., Almeida, A., Martins, C., Gonc¸alves, R., and

Martins, J. (2019). Using poi functionality and acces-

sibility levels for delivering personalized tourism rec-

ommendations. Computers, Environment and Urban

Systems, 77:101173.

Shen, J., Deng, C., and Gao, X. (2016). Attraction recom-

mendation: Towards personalized tourism via collec-

tive intelligence. Neurocomputing, 173:789–798.

Sun, X., Huang, Z., Peng, X., Chen, Y., and Liu, Y. (2019).

Building a model-based personalised recommenda-

tion approach for tourist attractions from geotagged

social media data. International Journal of Digital

Earth, 12(6):661–678.

Tintarev, N., Flores, A., and Amatriain, X. (2010). Off

the beaten track: a mobile ﬁeld study exploring the

long tail of tourist recommendations. In Intl. Conf. on

Human computer interaction with mobile devices and

services, pages 209–218.

Yochum, P., Chang, L., Gu, T., and Zhu, M. (2020). Linked

open data in location-based recommendation system

on tourism domain: A survey. IEEE Access, 8:16409–

16439.

ICSOFT 2021 - 16th International Conference on Software Technologies

372