Tourist Attraction Recommendation Service: An Approach,

Architecture and Case Study

Alexander Smirnov

1,2

, Andrew Ponomarev

1,2

and Alexey Kashevnik

1,2

St. Petersburg’s Institute for Informatics and Automation of the RAS, 14

Line, 39, St. Peterburg’s, Russian Federation

ITMO University, Kronverksky, 49, St. Peterburg’s, Russian Federation

Keywords: Recommendation System, Tourism, Context-awareness, Collaborative Filtering, Point-of-interest,

Knowledge, Ontology.

Abstract: The paper proposes a complex multi-model approach to recommendation systems design in the domain of

tourist information support. Specifically, it proposes to construct a recommendation system as a

composition of loosely coupled modules, implementing both personalized and non-personalized methods of

recommendations and a coordination module responsible for adaptation of the whole system to the specific

tourist and situation context.

1 INTRODUCTION

Tourism has become one of the largest and fastest-

growing economic sectors in the world. Despite

occasional shocks, it has shown virtually

uninterrupted growth. International tourist arrivals

have increased from 527 million in 1995 to 1133

million in 2014. International tourism receipts

earned by destinations worldwide have surged from

US$ 415 billion in 1995 to US$ 1245 billion in 2014

(World Tourism Organization, 2015). Moreover, the

number of international tourist arrivals worldwide is

expected to increase by an average of 3.3% a year

over the period 2010 to 2030 (ibid.).

At the same time, there are some structural and

behavioral changes in tourism highly connected to

the development of Internet and Information

Technologies. The increasing use of ICTs in tourism

services allows tourists to take a more active role in

the production of tourism products, being no longer

satisfied with standardized products. The

”postmodern tourist”' with differentiated life-styles,

individual motives and specific interests demands

products tailored accordingly to stated preferences

(Berka, T. and Plößnig, M., 2004).

All that makes the problem of tourists’

information support more actual than ever.

Therefore, information (and search) services of all

kinds that can help in collecting information about

the trip being planned and provide tourist with

information needed during the trip are becoming

more and more popular. One of the functions

typically provided by those services is

recommendation of attractions based on tourist’s

preferences and current conditions (weather,

transport, etc.).

Systems intended to mitigate a choice problem

leveraging (implicit or explicit) subjective

preferences received a name of “recommendation

systems”. The variety of techniques to build, deploy

and assess this kind of systems separated into a

specific research area in the mid-90s of XX century.

Approaches to build recommendation systems

are usually classified according to the kind of input

data that is used for recommendations. Most popular

are two of them (Adomavicius G. and Tuzhilin A.,

2005): collaborative filtering and content-based. In

the former one the only information that is available

are ratings that users assigns to objects. In the latter,

input information is formed by structured

representation of items and a vector of user’s ratings.

There are several more approaches: demographic

recommendation systems, knowledge-based

recommendation systems, social-based

recommendation systems, but they are less used.

Dependence on a specific type of information

causes limitations in applying each of

recommendation techniques. For example,

collaborative filtering cannot be used when the

number of ratings is small, but just after start of any

recommendation system the set of ratings is usually

empty, hence the so called “cold start” problem.

Smirnov, A., Ponomarev, A. and Kashevnik, A.

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study.

In Proceedings of the 18th International Conference on Enterprise Information Systems (ICEIS 2016) - Volume 2, pages 251-261

ISBN: 978-989-758-187-8

251

Similarly, the structured representation of items

needed for content-based methods might be

available for one regions (in tourist recommender

systems) and be missing for others.

This paper proposes multi-model approach,

which prescribes creation of a family of

recommendation modules, based on various

principles and relying on different types of input

information. Coordination of the modules and

composing an aggregate recommendations list,

taking into account current situation, is performed

by a coordination module. Employed basic models

range from knowledge-based fuzzy rules to

collaborative filtering and coordination module

leverages fuzzy inference to evaluate each basic

model in the current context.

There is a prototype implementation of the

recommendation service, functioning as a part of

tourist information support system «TAIS – Mobile

Tourist Guide” (Smirnov et al., 2014).

2 RECOMMENDATION

SYSTEMS IN TOURIST

INFORMATION SUPPORT

Key premises to employ recommendation systems in

some domain are, first, abundance of choice

complicating decision-making, and second,

significant subjectivism in decision-making. Tourist

in an unfamiliar environment (city, country)

frequently face both of the premises: abundance of

attractions to visit and uncertainty in which of them

to visit to gain most positive experience from the

stay. This explains the attention that is paid to tourist

information support in recommendation systems

community. Besides, social sciences research reveals

the importance of decision support systems in

tourism, caused by large number of aspects that need

to be paid attention to: tourist mobility, high risk and

uncertainty in unfamiliar environment, distributed

nature of information sources and several other

factors (Gretzel, 2011).

Main directions and achievements in tourist

recommendation systems design are summarized in

review papers (Kabassi, 2010) (systems before fall

2009) and (Borràs et al., 2014) (2008-2014). These

studies reveal that nowadays in tourist

recommendation systems all modern

recommendation techniques are used. Collaborative

filtering, content-based and demographic ones are

the most widely employed.

There is also another branch of research that may

be relevant to tourist recommendation system

community. It is media stream analysis, employed

e.g. for point-of-interest (POI) detection (see, e.g.,

Han and Lee, 2015). In some sense, this technique

can be interpreted as an “open form” of

collaborative filtering. Openness here means that

this technique does not implement functionality of

user’s feedback collection typically present in

recommendation systems; instead, it relies on some

external feedback source, namely, social media

event stream. A common idea of this kind of

systems is that geo-tagged images are interpreted as

some signs that a user who posted them enjoyed the

place or view. Additional analysis can help make

further inference, e.g., Han and Lee (2015) try to

distinguish images made by tourists (which are

mostly relevant for making recommendations to

tourists) from images made by local population by

analysing attributes of the image poster’s account in

Flickr (the primary media event stream for that

system). In (Crandall et al., 2009) city attractions are

visualized based on the analysis of images from

photo sharing service. In (Marcus et al., 2011)

methods are proposed to detect actual events taking

place in city based on the Twitter stream.

Photo2Trip system (Lu et al., 2010; Yin et al., 2010)

makes step further; based on the analysis of

sequences of geo-tagged photos from public photo

sharing sites, Photo2trip identifies and recommends

typical tourist trips.

To the best of authors’ knowledge, there are no

systems trying to integrate “classical”

recommendation approaches (e.g., content-based,

collaborative filtering) with emerging “open form”

non-personalized recommendation techniques based

on the social media event stream analysis.

3 MULTI-MODEL APPROACH

FOR RECOMMENDATION

SYSTEMS DESIGN

As it was noted earlier, there is a set of well-known

approaches to make recommendations. Main

criterion for distinguishing between them is the kind

of information used in the respective approach. Each

existing approach inevitably bears some advantages

and disadvantages. In the recommendation systems

research specific consequences of disadvantages

have received metaphorical (usually) names: “cold

start problem”, “grey sheep problem”. The former

one refers to impossibility of a recommendation

system that is based solely on historical data, to

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

252

make recommendations to new users (without any

historical data associated with them) or recommend

new items (not rated by any user yet). The latter one

refers to difficulties in dealing with non-typical

users. These problems on a higher level are

consequences of incompleteness of input data and

assumptions that are immanent to recommendation

systems.

As each of the “pure” approaches to making

recommendations is based on its own set of input

data, it is natural to compensate paucity of

information of one type (more difficult to obtain) by

information of some other type, leveraging several

“pure” approaches simultaneously. This is how

hybrid recommendation systems work.

Hybridization may touch different levels of the

system. For example, one of the most popular forms

of hybridization is “collaboration via content”

(Blanco-Fernández et al., 2008), which is based on

collaborative filtering, but similarity measure

between users is modified in such a way that it

considers not only ratings, but also similarity of

some semantic attributes of users. “Collaboration via

content” is an example of “deep” hybridization as it

transforms the algorithm of one of “pure”

approaches, adding to it some new information. An

example of “shallow” hybridization is an ensemble

of recommendation systems that work independently

with their results merged.

In this paper, recommendation system is built by

similar “shallow” hybridization. The system

includes several independent modules, each of them

implementing one recommendation algorithm

(mostly, “pure” ones, as they are more tried-and-

tested). Along with recommendation modules, the

system includes coordination module that merges

recommendations generated by “pure” algorithms,

using knowledge about their strong and weak sides

and current situation. For example, if there are not so

many ratings in the database, then recommendations

of collaborative filtering module will likely be

regarded as not reliable. Advantage of this modular

construction is the simplicity of adding new

recommendation modules – it requires mapping its

input specification on the information model of the

system and (in some cases) modifying the

coordination module. Coordination module is

intentionally designed as configurable.

Recommendation service provides POI

recommendations on two levels: non-personalized

and personalized. Personalized recommendations are

usually preferable, but their generation requires

information about user-item interactions. If this

information is not available, the service falls back to

non-personalized recommendations, that requires

only aggregate information about item popularity,

For example, if there is no information about users’

preferences, then upon trip to St. Petersburg it is

only possible to recommend popular tourist

locations like Hermitage museum, Peter and Paul

fortress, or St. Isaac’s Cathedral, and these

recommendations can be made only on the basis of

the statistics of visiting. If, however, the user stated

in her profile that she is interested in engineering,

and it is known that she visited Centre Pompidou

during her trip to Paris and enjoyed it, then it is

possible to recommend Central Railway Museum,

Central Museum of Communications, or Erarta

Museum of Contemporary Art.

However, on both levels POI recommendation

may, and usually should be context-aware. In the

non-personalized recommendations level context-

awareness stands for using a stratified and time-

bound data for making recommendations. Non-

context-aware non-personalized recommendations

would be based on overall visiting statistics

(actually, for all time). The simplest form of context-

awareness in this case, would be making

recommendations on yearly statistics data, which

would help to identify and recommend places that

are most popular now recently. The more elaborate

form is making stratified samples, attributing visitors

to days of week, months of year, country of origin

etc. On the other hand, non-context-aware

personalized recommendations are well-known

classical collaborative filtering in the space of solely

user-item ratings, or content-based methods

matching users to some features of items. Context-

awareness would usually mean attributing each

rating set by the user to some kind of external

conditions and limiting ratings used for

recommendations to those, which are attributed to

similar conditions.

3.1 Non-personalized

Recommendations

For non-personalized recommendations that are

most actual in the paucity of preferences data, the

TAIS’s attractions recommendation service

leverages the publicly available geo-tagged stream

of events (photos and tweets). As it was discussed

earlier, non-personalized recommendations are

based on visiting statistics data. There are three

potential sources of these data: a) the data can be

collected by the tourist application itself (TAIS in

this case); b) the data can be queried for from be

local authorities or POI administration; c) the data

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study

253

can be mined from the global stream of public data.

The source (a) is the most convenient as the data can

be collected with all the needed context attributes

and in the most appropriate form and granularity,

however, it requires a huge number of users and

cannot be employed by a newly created application.

The source (b) relies on the communication with

external entities (local authorities and museums

administration) and is very laborious. It can be

appropriate for a local application, e.g.

St. Petersburg local city guide, but hardly can be

implemented for a global recommendation service

that should work in every location worldwide.

Moreover, it is not suitable for recommending

architectural POIs, publicly available observation

places, as there might be no administration to collect

visiting statistics. With all the drawbacks of (a) and

(b) for making a globally active POI

recommendation service, the option (c) becomes

viable. With the dissemination of camera and GPS-

equipped mobile devices, widening mobile internet

coverage and the forming of new information

processing habits, publicly available stream of geo-

tagged events is becoming more and more affluent.

There are many scientific publications showing

various ways of leveraging this live source of human

activity: from events and opinion detection to, a

more relevant to the topic of this paper, POI

detection and recommendation. However, this

stream obviously bear some bias that must be taken

into account. E.g., it is produced by active users of

social networks and owners of modern smart-

phones. Target users of mobile tourist guide are

obviously a subset of smart-phone owners, but

actually not necessarily are active users of social

networks, so there still is a chance of biased

inference.

In the working prototype of the recommendation

service Flickr is used as a source of geo-tagged

photos. Each geo-tagged photo is interpreted as an

evidence that some user has visited certain

geographical location. In contrast with classical

recommendation systems, making photo doesn’t

express explicitly user’s attitude to the object in the

frame. So, not necessarily making photo is

equivalent to marking the place with “like” or

setting it a good rating. However, as the research on

geo-tagged social media reveals, geo-tags are

concentrated around attractive landmarks and can be

used to detect them.

As TAIS application is targeted on the support of

tourists during trip, it is important to differentiate

between photos that are likely to be done during trip

from those that are done, for example, at the work,

after work and even during weekend by local

population. In this work the approach described in

(Han and Lee, 2015) is adopted. It is based on the

analysis of Flickr user’s profile, specifically its

location attribute. This attribute value designates a

location that is perceived as home for a user.

Therefore, all the photos that are not associated

(taken farther than 30km) with this location are

considered to be done on trip and are processed by

the procedure of popular tourist places detection.

Therefore, a context of a photo for the non-

personalized recommendation includes:

- date and local time when photo was taken with

derivate values;

- photo poster’s home location.

As it was discussed earlier, photo poster’s home

location is used to select relevant photos, whereas

date and time are used to build spatio-temporal map

of location popularity.

All the area is split on rectangles approximately

100x100 meters (10

-4

degrees latitude and variable

range in longitude, depending on the latitude), then

the number of tourist photos in each rectangle is

counted, taking into consideration local time the

photo was taken and its derivative attributes: day of

week, time of day and season. This number is

attributed to all POIs that are located in the

rectangle.

One impediment caused by the way Flickr API is

organized, is that it is relatively easy to estimate the

number of photos in the area, but getting the

attributes (shot time, date and precise location,

poster’s home location) of all the photos is

impractically slow due to the number of API calls

required. To solve this problem, sampling technique

is used.

Figure 1: Example statistics on the number of photos made

by Flickr users in St.Petersburg in 2015.

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

254

3.2 Personalized Recommendations

Personalized recommendations account for user’s

preferences and potentially are more accurate,

however, they demand more information.

Particularly, user preferences should be defined in

some form. In the proposed service personalized

recommendations are represented by context-aware

collaborative filtering.

As TAIS application is targeted on the support of

tourists during trip, it is important to differentiate

between photos that are likely to be done during trip

from those that are done, for example, at the work,

after work and even during

3.2.1 Context-aware Collaborative Filtering

One of the promising directions to improve the

predictive quality of recommendation systems in

general (and collaborative filtering systems among

them) is context-awareness (Adomavicius et al.,

2011). The context describes conditions in which the

user rates an object or asks for recommendations.

In the proposed tourist attraction information

service the following context attributes are

distinguished:

a) time;

b) company in which the user visited the

attraction (alone, with a friend or with the family);

c) weather (sunny, rainy, etc).

Values are assigned to these attributes in mostly

automated fashion. For example, the user opens the

attraction evaluation screen being near to that

particular attraction (according to the mobile

device’s GPS sensor). In this case the time attribute

is filled in with the current time and current weather

is queried from the context service. However, there

is also a possibility to set the values of context

attributes manually in the evaluation screen of the

mobile application. It is convenient, for example, if a

user wants to rate the attractions seen during the day

upon returning to the hotel in the evening. To

facilitate deferred evaluation the proposed system

tracks attractions the user visits and shows unrated

visited attractions in a special screen. The user does

not have to assign values to each context attribute. If

a context attribute is not given a value, it is assumed

to have value “any”.

There are three general approaches to take

context into account in recommendation systems

(Adomavicius et al., 2011): (a) contextual pre-

filtering; (b) contextual post-filtering; (c) contextual

modelling.

The advantage of the contextual pre-filtering and

post-filtering approaches is that they are compatible

with classical (not context-aware) recommendation

algorithms. The context awareness in these

approaches comes true by transformation of either

input or output of the classical recommendation

algorithm. In the contextual pre-filtering approach,

the rating data that is not related to the context is

filtered out before applying the recommendation

algorithm. On the other hand, in the contextual post-

filtering approach the resulting list of

recommendations is ordered or filtered taking into

account context values.

In the contextual pre-filtering approach all the

ratings that are irrelevant to the context discarded

from the rating matrix before the recommendation

algorithm is applied. For example, if in some

attraction recommendation service the context

includes weather conditions, then making

recommendations in a rainy day should not use

ratings assigned in sunny days. This approach

aggravates the important problem inherent to

collaborative filtering systems – rating matrix

sparsity. The main goal pursued by contextual pre-

filtering methods is to take into account the context,

but not let rating matrix to become too sparse.

In the proposed system the context generalization

method (Adomavicius et al., 2005) (one of the

contextual pre-filtering methods) is used for taking

context into account. In this method, the rating

matrix is filtered not only by exact values of context

attributes, but also by its possible generalizations. To

use this method the context model has to support

context generalization. In most general form, it

means that at least one context attribute must be

defined on a set with a strict partial order relation of

generalization (→). Let A be a set of attribute values

and a

, a

∈ A. Then notation a

→ a

means that

value a

is a generalization of a

. A context is usually

represented by m attributes. Let c = (c

, …, c

) and

c’ = (c’

, …, c’

) are two contexts. We define с’ as a

generalization of c (c → c’), iff there exists at least

one i ∈ {1,..,m}, such that c

→ c’

. We call context с

incompatible with c’, iff neither c → c’ nor c = c’. In

most cases, the generalization relation forms some

kind of a hierarchy (or multiple hierarchies).

In the proposed system, the context

generalization is enabled by following:

a) The set of Time attribute values includes not

only exact date and time values but also “any” value

and aggregate values for each season, day type

(working day or weekend) and time of day

(morning, afternoon, evening). The generalization

relation is defined naturally.

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study

255

b) The set of Company attribute values includes

values “alone”, “with friends”, “with family” and

“any”. “Any” value is defined to be a generalization

of any other value.

c) The set of Weather attribute values includes

values “sunny”, “rainy”, “cloudy”, “snowy” and

“any”. “Any” value like in (b) is defined to be a

generalization of any other value.

For example, the exact context could be (Time:

“July 31, 2013 17:30”; Company: “with family”;

Weather: “sunny”). This context can be generalized

to (Time: “summer”; Company: “with family”;

Weather: “sunny”) or even to (Time: “summer”;

Company: “any”; Weather: “any”).

It is obvious that a context can be generalized in

several ways and directions. In systems with many

attributes and many levels of granularity of

attributes, enumerating all possible context

generalizations is a problem and various heuristics

are used for picking appropriate generalizations

(Adomavicius et al., 2005). In the proposed system,

there are not so many possible generalizations, so all

of them are enumerated through implicit directed

graph traversal procedure. The nodes of this graph

are attribute values and the arcs are generalization

relations.

A user rates attractions on a five-point scale (1 –

bad, 5 – excellent). The rating obtained from the

user (raw rating) is normalized to reduce individual

bias in assessment: some users tend to put relatively

high ratings to all attractions, others in contrary tend

to put relatively low ratings. Normalized rating

given by user u to attraction j is defined by formula:

⎟

⎠

⎞

⎜

⎝

⎛

−=

∑

∈

ujuj

rr 3

here, r

is raw rating of the attraction j given by

user u, and K

is a set of all attractions rated by user

u. The idea of normalization is to shift from user-

oriented five-point scale to calculations-oriented

zero-centered scale. The sign of the normalized

rating corresponds to general attitude of the user

(whether it is positive or negative) and the absolute

value of the rating corresponds to the strength of that

attitude. The straightforward way to normalize

ratings is to subtract scale average (i.e. “3”) from

each rating. It would work nice if users normally

used all the range of five-point scale. However, most

users in fact rate items using some subset of the

scale, e.g., only “3”, “4” and “5”. In this case

subtracting scale average would result in non-

negative normalized ratings missing the fact that the

user definitely likes items he/her rated “5” and

probably doesn’t like items rated “3”. Hence, the

normalization procedure should capture not only the

scale characteristics but also the observed usage of

this scale. Therefore, a popular method of

normalization is subtracting average user rating from

all his/her ratings. This method works well in most

cases but have some subtle drawback which turns

out when there are only a few ratings. For example,

when the user rated only two items – both with “5” –

then normalization over the average user rating

would turn these ratings into zeroes. I.e. a priori

notion of five-point scale with “5” as the best mark

is lost in favor of adaptation to the observed usage of

this scale. To alleviate this drawback in the proposed

system we use slightly modified version of the

normalization over the average user rating. During

the normalization we add one fake rating of “3”

(scale average) to the set of user ratings having a

purpose to stick other ratings to the original notion

of the scale. This modification is significant when

there are a few ratings (in the example above two

“5” ratings become positive) but its contribution to

the normalized ratings vanishes as the number of

users’ ratings grows.

Attraction rating estimation for a given user is

performed in two steps:

1) a group of users with ratings similar to the

given user’s is determined;

2) rating of attraction is estimated based on

ratings of this attraction assigned by users of the

group.

While building the list of recommendations,

several possible generalizations of the context is

used. For each context generalization ratings

received in contexts incompatible with this

generalization are not taken in to account.

User group is determined by k-Nearest

Neighbours method (kNN). The similarity between

users u and v is calculated as a cosine measure

between normalized ratings vectors of users

according to the following formula:

∑∑

∑

∈∈

∈

uvuv

viui

Here I

is a set of attractions rated by both users

u and v.

Attraction rating estimation for the user is based

on ratings of that attraction assigned by other users

of the group with respect to their similarity to the

user. It is calculated as a weighted average of

normalized ratings among group members:

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

256

∑

∈

uvvj

here G is the group of the user.

3.2.2 Context-aware Knowledge-based

Recommendations

This recommendations module uses the attractions

data extracted from open internet services, tourist

type and context data. It is driven by a knowledge

base connecting tourist properties, attraction

properties and context parameters. The advantage of

this approach is that this module does not require

ratings history and therefore, it can be used

immediately after recommendation service

deployment.

Problems of tourist industry development have

received much attention in the scientific literature.

From the point of this paper, the most valuable are

attempts to build a tourist typology and link different

types of tourists to their preferred types of activities

during trip. One of the first papers proposing a

typology like that was (Cohen, 1972), with 4 types

of tourists. Later, other typologies either for all the

variety of tourists (Wickens, 2002; Gibson and

Yiannakis, 2002), or for some subset of them were

proposed (Mehmetoglu, 2007; McKercher and Du

Curos, 2003).

The knowledge-based recommendation module

uses typologies, proposed in (Gibson and Yiannakis,

2002; McKercher and Du Curos, 2003), as Gibson

and Yiannakis (2002) propose a typology with a

greater differentiation (15 roles), which allows to

specify preferences of each role more precisely.

McKercher and Du Curos (2003) propose a cultural

tourist typology, and cultural tourism is one of the

focuses of this paper.

To fill the knowledge base the results of several

scientific publications (Pearce and Packer, 2013;

Hannam et al., 2014; Park et el., 2012) were used

linking types of tourists and their preferred activities.

Fuzzy logic is used to represent the properties of the

tourist and of the situation. It is caused by the fact that

crisp classifications rarely can be applied to cultural

objects or people (Gibson and Yiannakis, 2002).

Linguistic variable is defined by a tuple (x, T, U,

G, M), where x is a name of the variable, T – term-

set, each element of which (a term) is represented as

a fuzzy set on the universe U; G – syntactical rules

of new terms construction, often in the form of a

grammar; M – semantic rules, defining membership

functions of fuzzy sets in T.

All the linguistic variables in the recommendation

module can be divided into three groups:

1) Variables that describe a tourist type,

according to (Gibson and Yiannakis, 2002;

McKercher and Du Curos, 2003). Their names are

synthesized as prefix “Type_” followed by a type

abbreviation (e.g., Type_SNL corresponds to “Sun

Lover” type from Gibson and Yiannakis (2002)).

Term-set for all these variables is a set T =

{“Definitely true”, “Likely true”, “Likely not true”,

“Definitely not true”}, and a universe – U = [0; 1].

2) Variable Weather, describing the weather in

fuzzy linguistic terms.

3) Output variable Recommend, having term-set

{“Definitely recommend”, “Recommend”, “Neither

recommend or not”, “Not recommend”, “Definitely

not recommend”}.

The recommendations are formed by a set of

fuzzy rules involving statements with linguistic

variables and crisp predicates using Mamdani-type

inference (Mamdani, 1974). For example:

IF (Type_EDT IS “Definitely true” OR

Type_PCT IS

“Definitely true”) AND

ObjectType IS Museum AND

FreeTime > 2 hours

THEN

Recommend IS “Definitely recommend”

3.3 Coordination Module

The goal of the coordination module is to merge

results obtained by various recommendation

modules based on their expected trustworthiness

(which is related to the availability of the

information crucial for the algorithm implemented

by that particular module). Major criteria taken into

account by the coordination module are:

- collaborative filtering module requires

significant amount of user-item ratings;

- to receive recommendations with the help of

collaborative filtering module , a user should express

her preferences by making several ratings;

- non-personalized recommendations module can

function only in the areas where social media stream

is fairly intensive.

Therefore, each recommendation module has its

own restrictions, and the coordination module assigns a

degree of belief to each of the modules, based on

evaluating these restrictions. Coordination module is

also based on fuzzy inference. Linguistic variables used

in its knowledge base can be divided into 4 groups:

1) User characteristics (a number of ratings set

by the user, the number of user’s “neighbours” in the

collaborative filtering module).

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study

257

2) Ratings database characteristics (RMSE –

Root mean squared error, expected error of rating

prediction by collaborative filtering module).

3) Context characteristics (photos in the region).

4) Output variable Belief.

Rule example:

IF RatingsByUser IS “Many” AND

NeighborsOfUser IS “Many” AND

RMSE IS NOT “High”

THEN

Belief IS “High”

Degree of belief to the recommendations

provided by a module is obtained by applying rules

to the module and defuzzification of the output

variable. After that, recommendations provided by

each of the modules are weighted according to the

degree of belief, ranked, and shown to the user.

Overall architecture of the recommendation service

is shown in Figure 2.

4 TAIS – MOBILE TOURIST

GUIDE

TAIS is a mobile travel guide application based on

Smart-M3 platform (Honkola et al., 2010)

implementing a smart space concept. That allows to

significantly simplify further development of the

system, to add information sources and services, and

to make the system highly scalable. The key idea of

this platform is that the formed smart space is

device, domain, and vendor independent. Smart-M3

assumes that devices and software entities can

publish their embedded information for other

devices and software entities through simple, shared

information brokers. Platform is open source and

accessible for download at Sourceforge.

Implementation of TAIS application has been

developed using Java KPI library. Mobile clients

have been implemented using Android Java

Development Kit. The application consists of a set

of services (Smirnov et al., 2014) that interact with

each other for providing the tourist

recommendations about attraction that is better to

see around. There are client application, attraction

information service, recommendation

service(described earlier in this paper), region

context service, ridesharing service (Smirnov et al.,

2012), and public transport service.

The main application screen is shown in Figure

3, left screenshot. The tourist can see images

extracted from accessible internet sources around,

Figure 2: Recommendation service architecture.

• Attraction lists

• Ratings

• Geographic

coverage with

event statistics

Mobile application

Non-personalized

recommendations module

Context-aware collaborative

filtering module

Coordination module

Context-aware knowledge-

based recommendations

module

Preferences

collector module

Context processing

module

Recommendation modules

• Collect preferences

• Collect context

• Show recommendations

Data storage

Social media data

collector

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

258

Figure 3: Tourist guide screenshots: main screen, context menu with actions.

Figure 4: Tourist guide screenshots: attraction details and route to the attraction.

clickable map with his/her location, context situation

(weather), and the best attractions around ranked by

the recommendation service. When the tourist click

to an attraction the following context menu is

opened (see Figure 3, right screenshot). The tourist

can see detailed information about the chosen

attraction (Figure 4, left screenshot), browse

attraction reaching path that is proposed by the

system route to an attraction (Figure 4, right

screenshot), and/or estimate it (Figure 5, left

screenshot).

Detailed information about attraction contains a

list of images that is associated with this attraction

and it description. This information is extracted by

the attraction information service from different

internet sources (e.g., Wikipedia, Wikivoyage, and

Panoramio are currently used).

The tourist has possibility to rate images using

the following options: “like image”, “dislike image”,

“this image is not applicable” to the attraction (see

Figure 4, left screenshot). Based on these ratings the

recommendation service will re-order images for

this or another tourist next time.

The tourist can browse the attraction reaching

path by choosing “Show on the map” item in context

menu (see Figure 4, right screenshot). The routing

service that is responsible for path finding uses

OpenStreetMap-based web mapping service (Teslya,

List of images

Estimation options

Attraction description

Location Destination attraction

Path

List of images

Clickable

Map

Contex

Recommended

attractions

Location

Location on

the ma

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study

259

2014). Routing service provides the tourist

possibility to build pedestrian path, find fellow

travelers who go to the same direction (Smirnov et

al., 2012), and find public transport to reach chosen

attraction.

Tourist can browse information about the best

attractions around presented by the mobile tourist

guide in the main screen and click button “More” to

see more attractions (see left screenshot in Figure 5).

The tourist can estimate if an attraction is interesting

or not by looking through the names and thumbnail

images. If he/she would like to get more

information, it is possible to open description

window (Figure 4, left screenshot). Tourist guide

also calculates distance to every attraction (see

Figure 5, left screenshot).

The tourist can rate an attraction if he/she likes

or dislikes it (see Figure 5, right screenshot). For this

purposes he/she specify the context (company and

weather) and assign rating using five stars scale.

Pressing “menu” button allows to search

information for worldwide attractions by choosing

another area (country, region, and city) and to access

the settings page of the mobile tourist guide

application. In the status bar, the tourist can search

for attractions worldwide.

5 CONCLUSIONS

The paper proposes multi-model approach for

recommendation services for tourist domain, which

aims on dealing with paucity of input information

for recommendations. Its main idea is to create a

family of recommending modules, based on various

principles and requiring different input. To merge

and reconcile results obtained from these modules, a

special coordination module is introduced. It is

designed to be highly configurable and uses

declarative knowledge representation.

The proposed approach is implemented in the

recommendation service of “TAIS – Mobile Tourist

Guide” system, developed under the EU program for

cross-border e-tourism framework in Oulu Region

and the Republic of Karelia (Development of Cross-

Border e-Tourism Framework for the Program

Region – Smart e-Tourism (European Community –

Karelia ENPI CBC 2007-2013 Program, 2012-2014

– project KA322)).

Further work is aimed on implementation of new

recommendation modules and fine-tuning the

knowledge base of the coordination module, based

on the actual usage data.

ACKNOWLEDGEMENTS

The research was partially supported by grants # 14-

07-00345, # 15-07-08092, # 16-07-00463 of the

Russian Foundation for Basic Research and

project 213 (program I.5P) of the Presidium of the

Russian Academy of Sciences. This work was also

partially financially supported by the Government of

the Russian Federation, Grant 074-U01.

Figure 5: Tourist guide screenshots: all attractions around and estimation attraction interface.

Attraction image and

descri

tion

Distance to

an attraction

The tourist context

ICEIS 2016 - 18th International Conference on Enterprise Information Systems

260

REFERENCES

Adomavicius, G., Mobasher, B., Ricci, F., Tuzhilin, A.,

2011. “Context-aware recommender systems”, AI

Magazine, vol. 32(3), 2011, pp. 67–80.

Adomavicius, G., Sankaranarayanan, R., Sen, S., Tuzhilin,

A., 2005. “Incorporating contextual information in

recommender systems using a multidimensional

approach”, ACM Trans. Inf. Syst. vol. 23(1), Jan.

2005, pp. 103-145.

Adomavicius, G., Tuzhilin, A., 2005. Toward the next

generation of recommender systems: a survey of the

state-of-the-art and possible extensions // IEEE

Transactions on Knowledge and Data Engineering.

June 2005. vol. 17, no. 6, P. 734--749.

Berka, T. and Plößnig, M., 2004. Designing recommender

systems for tourism. In ENTER. Cairo. Available at:

http://195.130.87.21:8080/dspace/handle/123456789/5

83 [Accessed October 25, 2014].

Blanco-Fernández Y., Pazos-Arias J., Gil-Solla A.,

Ramos-Cabrer M., López-Nores M., García-Duque J.,

Fernández-Vilas A., Díaz-Redondo R., 2008.

Exploiting synergies between semantic reasoning and

personalization strategies in intelligent recommender

systems: A case study // The Journal of Systems and

Software, 81, 2008, 2371-2385.

Borràs J., Moreno A., Valls A., 2014. Intelligent tourism

recommender systems: A survey // Expert Syst. Appl.

2014. Т. 41. С. 7370–7389.

Cohen, E., 1972. Toward a sociology of international

tourism. Social Research 39, 1972, 164–172.

Crandall, D.J., Backstrom, L., Huttenlocher, D.,

Kleinberg, J., 2009. Mapping the world’s photos, in:

Proceedings of the 18th International Conference on

World wide web, WWW’09, ACM, New York, NY,

USA, ISBN: 978-1-60558-487-4, 2009, pp. 761–770.

http://dx.doi.org/10.1145/1526709.1526812.

Gibson, H., Yiannakis, A., 2002. Tourist Roles: Needs and

Lifecourse // Annals of Tourism Research, 2002, Vol.

29, No. 2, pp. 358-383.

Gretzel, U., 2011. Intelligent systems in tourism: A social

science perspective. Annals of Tourism Research,

38(3), 2011, 757–779.

Han J., Lee H., 2015. Adaptive landmark recommendations

for travel planning: Personalizing and clustering

landmarks using geo-tagged social media // Pervasive

Mob. Comput. 2015. Т. 18. С. 4–17.

Hannam, K., Butler, G., Paris, C.M., 2014. Developments

and key issues in tourism mobilities // Ann. Tour. Res.

2014. Т. 44. С. 171–185.

Honkola J., Laine H., Brown R., Tyrkko O., 2010. Smart-

M3 Information Sharing Platform // Proc. of ISCC

2010, IEEE Comp. Soc. Jun. 2010. Pp. 1041-1046.

Lu, X., Wang, C., Yang, J.-M., Pang, Y., Zhang, L., 2010.

Photo2trip: generating travel routes from geo-tagged

photos for trip planning, in: Proceedings of the

International Conference on Multimedia, MM’10,

ACM, New York, NY, USA, ISBN: 978-1-60558-933-6,

2010, pp. 143–152. http://dx.doi.org/10.1145/1873951

.1873972.

Mamdani, E. H., 1974. Application of fuzzy algorithms

for the control of a simple dynamic plant // In Proc

IEEE (1974), 121-159.

Marcus, M.S. Bernstein, O. Badar, D.R. Karger, S.

Madden, R.C. Miller, 2011. Twitinfo: aggregating and

visualizing microblogs for event exploration, in:

Proceedings of the 2011 Annual Conference on

Human factors in computing systems, CHI’11, ACM,

New York, NY, USA, ISBN: 978-1-4503-0228-9, 2011,

pp. 227–236.

McKercher, B., Du Cros, H., 2003. Testing a cultural

tourism typology // International Journal of Tourism

Research, 2003, 5(1), 45–58.

Mehmetoglu, M., 2007. Typologising Nature-Based

Toursits by Activity – Theoretical and Practical

Implications // Tourism Management, 2007, 28, pp.

651-660.

Park, H., Yoon, A., Kwon, H., 2012. Task Model and Task

Ontology for Intelligent Tourist Information Service //

2012. Т. 5. № 2. С. 43–58.

Pearce, P.L., Packer, J., 2013. Minds On The Move: New

Links From Psychology To Tourism // Ann. Tour. Res.

2013. Т. 40. С. 386–411.

Smirnov A., Kashevnik A., Ponomarev A., Shilov N.,

Shchekotov M., Teslya N., 2014. Smart Space-Based

Intelligent Mobile Tourist Guide: Service-Based

Implementation, Proceedings of the 15th Conference

of Open Innovations Association FRUCT,

St.Petersburg, Russia, ITMO university publisher

house, 21-25 April 2014, 126-134.

Smirnov A., Shilov N., Kashevnik A., Teslya N., 2012.

Smart Logistic Service for Dynamic Ridesharing,

Internet of Things, Smart Spaces, and Next Generation

Networking, 12th International Conference, NEW2AN

2012, and 5th Conference, ruSMART 2012, St.

Petersburg, Russia, August 27-29, 2012, pp. 140-151.

Teslya N., 2014. Web Mapping Service for Mobile Tourist

Guide, Proceedings of the 15th Conference of Open

Innovations Association FRUCT, Saint-Petersburg,

Russia, ITMO university publisher house, 21-25 April

2014, pp. 135-143.

Wickens, E., 2002. The Sacred and the Profane: a Tourist

Typology // Annals of Tourism Research, 2002, Vol.

29, No. 3, pp. 834-851.

World Tourism Organization, 2015. UNWTO Tourism

Highlights 2015 Edition, URL: http://www.e-

unwto.org/doi/pdf/10.18111/9789284416899

[Accessed December 10, 2015]

Yin, H., Lu, X., Wang, C., Yu, N., Zhang, L., 2010.

Photo2trip: an interactive trip planning system based

on geo-tagged photos, in: Proceedings of the

International Conference on Multimedia, MM’10,

ACM, New York, NY, USA, ISBN: 978-1-60558-933-6,

2010,

pp. 1579–1582.

Tourist Attraction Recommendation Service: An Approach, Architecture and Case Study

261