Journal Pre-proofs

Tourism Recommendation System Based on Semantic Clustering and Senti‐

ment Analysis

Zahra Abbasi-Moud, Hamed Vahdat-Nejad, Javad Sadri

PII: S0957-4174(20)31017-4
DOI: https://doi.org/10.1016/j.eswa.2020.114324
Reference: ESWA 114324

To appear in: Expert Systems with Applications

Received Date: 11 May 2020

Revised Date: 29 September 2020
Accepted Date: 13 November 2020

Please cite this article as: Abbasi-Moud, Z., Vahdat-Nejad, H., Sadri, J., Tourism Recommendation System
Based on Semantic Clustering and Sentiment Analysis, Expert Systems with Applications (2020), doi: https://
doi.org/10.1016/j.eswa.2020.114324

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover
page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version
will undergo additional copyediting, typesetting and review before it is published in its final form, but we are
providing this version to give early visibility of the article. Please note that, during the production process, errors
may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2020 Elsevier Ltd. All rights reserved.

Tourism Recommendation System Based on Semantic Clustering and
Sentiment Analysis

Zahra Abbasi-Mouda , Hamed Vahdat-Nejada*, Javad Sadrib

aFaculty of Electrical and Computer Engineering, University of Birjand, Birjand, Iran
[email protected], [email protected]
*Corresponding author

b Computer Science and Software Engineering Department, Concordia University, Montreal, Quebec, Canada
[email protected]
ABSTRACT
Numerous number of tourism attractions along with a huge amount of information about them on web and social platforms have made the decision-
making process for selecting and visiting them complicated. In this regard, the tourism recommendation systems have become interesting for
tourists, but challenging for designers because they should be able to provide personalized services. This paper introduces a tourism
recommendation system that extracts users’ preferences in order to provide personalized recommendations. To this end, users reviews on tourism
social networks are used as a rich source of information to extract their preferences. Then, the comments are preprocessed, semantically clustered,
and sentimentally analyzed to detect a tourist’s preferences. Similarly, all users aggregated reviews about an attraction are utilized to extract the
features of these points of interest. Finally, the proposed recommendation system, semantically compares the preferences of a user with the features
of attractions to suggest the most matching points of interest to the user. In addition, the system utilizes the vital contextual information of time,
location, and weather to filter unsuitable items and increase the quality of suggestions regarding the current situation. The proposed recommendation
system is developed by Python and evaluated on a dataset gathered from TripAdvisor platform. The evaluation results show that the proposed
system improves the f-measure criterion in comparison with the previous systems.

Keywords: Tourism recommendation system, Sentiment analysis, Context-awareness, Semantic similarity

1. Introduction
With the expansion of tourism websites and social networks, a huge amount of data and comments are
produced and posted regularly (Ghane’i-Ostad, Vahdat-Nejad, & Abdolrazzagh-Nezhad, 2018; Neidhardt,
Rümmele, & Werthner, 2017) .People who are planning for a trip use this data as well as reviews of other
tourists as rich sources of information to select their appropriate destinations and points of interest (Hayashi
& Yoshida, 2018; Renjith, Sreekumar, & Jathavedan, 2020). However, it is a major challenge for tourists
to manually process large volumes of data (Borràs, Moreno, & Valls, 2014). In this regard, various tourism
recommendation systems have been proposed that try to provide personalized suggestions to users. They
aim to extract user preferences and present recommendations that are more in line with their preferences
(Abel, Herder, Houben, Henze, & Krause, 2013). Some recommendation systems cluster users based on
similarity in previously visited places and provide the same recommendations to users of each cluster
(Esmaeili, Mardani, Golpayegani, & Madar, 2020; Wan, Hong, Huang, Peng, & Li, 2018). For example, in
a recommendation system, the same paths are suggested to the users with similar profiles (Alrasheed,
Alzeer, Alhowimel, & Althyabi, 2020).
It is noteworthy that the users’ visit to a tourist attraction by itself does not provide enough information and
their reviews regarding these places are also important. As a result, another set of recommendation systems
leverage an analysis of comments to extract user preferences (Xiang, Du, Ma, & Fan, 2017). In this regard,
user reviews are analyzed and compared with attractions’ metadata, and places with the highest matching
are suggested (Leal, González–Vélez, Malheiro, & Burguillo, 2017). In this type of tourism
recommendation systems, the frequent keywords used in the comments are exploited, regardless of the
sentiments of users. As a result, the negative words that are stressed in user’s text might be mistakenly
returned as their preferences. While sentiment analysis is important in the tourism domain, it has been
overlooked in most cases (Alaei, Becken, & Stantic, 2019). In fact, a comprehensive recommendation
system in the context of tourism should include the following features:
 It should identify preferences by looking for concepts instead of being limited to specific keywords.
 It should leverage sentiment analysis on user comments to identify their positive versus negative
preferences.
 It should provide context-aware recommendations (Vahdat-Nejad, 2014), which are adapted to
user’s current situation.

To the best of the authors’ knowledge, none of the available tourism recommendation systems meets all of
the above features. Hence, this paper proposes a context-aware tourism recommendation system by
sentiment analysis. In this system, text processing and sentiment analysis are leveraged to extract users’
preferences, precisely. The preference extraction part is an extension to the initial paper (Abbasi-Moud,
Vahdat-Nejad, & Mansoor, 2019) in which user reviews about various attractions are extracted and
preprocessed. Then the preferences are extracted through semantic clustering and sentiment analysis. This
research extends the initial idea by extracting features of attractions from aggregated users’ reviews and
proposing a personalized recommendation system. The system additionally utilizes user contextual
information including location (in order to identify the attractions around him/her), time (to check when the
attractions could be visited), and weather (to provide recommendations that are appropriate to the current
weather situation). The proposed system is developed by Python. Therefore, an experiment is conducted
on TripAdvisor1, as a well-known travel platform, to evaluate the proposed system. In this regard, a dataset
including 100 users comments and visits in 2018 has been gathered. The evaluation results reveal high
efficiency of the proposed recommendation system in terms of precision, recall, and f-measure.
The structure of the paper is as follows. In the next section, a review of the related literature is provided.
Section 2 introduces the proposed recommendation system. In the section 4, details of the implementation
and evaluation, and in the final section, the conclusion remarks are presented.
2. Related work
Most of the tourism recommendation systems are based on geographic tags(Cai, Lee, & Lee, 2018; Lyu,
Chen, Xu, & Yu, 2020). As an example, the tags of photos shared on Flickr2 are used to identify and cluster
users who follow similar paths and this information is leveraged in order to recommend tourist
attractions(Majid et al., 2013). Due to the importance of user reviews about the places they have visited, a
number of recommendation systems utilize comments analysis. Moreover, context-awareness plays a major

1 www.TripAdvisor.com
2 www.Flickr.com
role in improving the quality of the tourist recommendation systems. Context-aware recommendation
systems are more successful in perceiving users’ preferences (Kulkarni & Rodd, 2020).
In this section, at first context-aware tourism recommender systems and then tourism recommender systems
that are based on user reviews are investigated, respectively.
Among many contextual information, location is the most important element used in current tourism
recommendation systems (Abowd et al., 1999; Yochum, Chang, Gu, & Zhu, 2020). In this regard, the
behavioral patterns of people who travel to the protected areas of the Ningaloo Marine Park in Australia
were extracted using GIS (Smallwood, Beckley, & Moore, 2012). Additionally, the user's current position
has been used to provide recommendations (Tumas & Ricci, 2009). These recommendations are displayed
based on user preferences for both public transportation and on foot. However, the major drawback of
methods that only use location as the context information is their domain constraints and their one-
dimensionality.
User's various contextual elements including location, speed, and route have been exploited to provide
personalized recommendations for visiting their favorite tourist destinations (Barranco, Noguera, Castro,
& Martínez, 2012). Furthermore, PSiS (Anacleto, Figueiredo, Almeida, & Novais, 2014) suggests the user
suitable destinations in accordance with their contextual information, including location, time, speed,
direction and weather. In this system, the users’ tourism history is used as their preferences to provide more
accurate recommendations. A similar tourism recommender system has been proposed for Chinese
language, which targets 100 prominent attractions of Taiwan (Yeh & Cheng, 2015). As individuals with
different occupations, ages and nationalities usually have different interests and perspectives, the age,
nationality and income of the user have been used as context elements to improve the accuracy of the
recommendations(Lu, Wu, Mao, Wang, & Zhang, 2015). Finally, TripAdvisor data is utilized to predict
users preferences (Pantano, Priporas, Stylos, & Dennis, 2019). In this system, each user should select at
least three topics from 18 tourism topics (e.g. Eco-tourism, Nature lover, etc.). Therefore, similar set of
items are suggested to the users with similar requirements and interests.
In another research, users’ preferences are derived based on the visiting time of different attractions
categories. Besides, unsupervised deep learning is exploited to detect the category of each attraction. The
system recommends attractions based on their similarity with the user’s preferences; however, it does not
consider users’ comments and feeling after their visit (Chen, Zhang, Cao, Wu, & Cao, 2020).
Most of the traditional tourism recommendation systems are based on user ratings on visited attractions.
For example, users’ preferences are extracted based on their ratings (1-5 points) on visited places. (Pu, Du,
Yu, & Feng, 2020). As text can bear much more information than a rating, reviewing users’ comments can
greatly improve the accuracy of these systems (Xiaoyao Zheng, Luo, Sun, Zhang, & Chen, 2018). Hence,
researchers have been analyzing user opinions by text-mining techniques (Bao, Fang, & Zhang, 2014;
Xiaolin Zheng, Ding, Lin, & Chen, 2016). In this regard, Loh et al. have designed a private chat page that
asks questions from users in order to find specific vocabulary within the scope of an ontology. Then the
type of attractions desired by the user is extracted based on the ontology concepts and recommendations
are provided, accordingly (Loh, Lorenzi, Saldaña, & Licthnow, 2003).
POST-VIA360 (Colomo-Palacios, García-Peñalvo, Stantchev, & Misra, 2017) is a bio-inspired
recommender system aims to make loyalty to the tourists after their first visit of an attraction. It uses a
tourism Ontology to provide suggestions based on previous visits, current location and social aspects.
Besides, Looker (Missaoui et al., 2019) creates a profile for each user based on their reviews. The main
drawback of this system is that it utilizes all words of users’ comments, which results in considering
invaluable words .
Yochum et al. have created a knowledge graph for Bangkok's tourist attractions (Yochum, Chang, Gu, Zhu,
& Zhang, 2018). This graph is based on the words in the text of the users’ comments, each of which is
considered as a concept. Then the characteristics of tourists and tourist attractions are shown as a concept
vector. Finally, using the cosine similarity measurement between vectors, the correlation between tourists
and attractions is calculated. Another study uses Topic modeling to extract titles from the user's reviews
and other tourists' reviews about nearby attractions (Leal et al., 2017). Then, using the semantic similarity
based on WordNet, the similarity of attractions is compared with the user's preferences and the most similar
attractions are recommended.
Considering users’ sentiments is vital in extracting their preferences from their reviews to avoid identifying
negative points as preferences. To this end, this paper proposes a tourism recommendation system that
exploits sentiment analysis and text mining to identify user preferences as well as key features of tourist
attractions. In fact, the proposed method augments previous research by considering sentiment analysis in
text mining on tourist’s reviews. Besides, in contrary to the reviewed papers, the proposed system is
context-aware, in which several contextual elements including time, location, weather condition, user’s
preferences, and attractions features are taken into account.
3. Proposed system
The proposed tourism recommender system consists of three stages. In the first stage, user preferences are
extracted from their comments and reviews. Similarly, in the second stage, the characteristics of tourist
attractions are extracted from the reviews of tourists regarding them. Finally, in the third stage, appropriate
recommendations are presented based on the contextual information as well as the similarity between user
preferences with the characteristics of the tourist attractions. The contextual information used in this method
includes weather, time, location, and user preferences. The pseudo code of the proposed recommender
system is presented in algorithm 1. The detail is discussed below.
 Algorithm1: The proposed recommendation algorithm

PoI Data: A = (<PoI a, Location l, Reviews 𝑟𝑎𝑙𝑙 𝑡𝑜𝑢𝑟𝑖𝑠𝑡𝑠 >)

Inputs:
Inputs: User Data: U = (<Location l, Reviews 𝑟𝑢𝑠𝑒𝑟 >)

Context Data: C = (<Location l, Weather w, Time t>)

Output: Ordered list of recommended PoIs per user: 𝑃𝑜𝐼𝑢 = [𝑃𝑜𝐼𝑎 , . . . , 𝑃𝑜𝐼𝑛 ]

Step1: Extract user preferences

In weather w
Step2: for each nearby user PoI do: 𝑇𝑜𝑝𝑃𝑜𝐼 Five most repetitious concepts of reviews of all
tourists

for each nearby user PoI do similarity-PoI=Compute Similarity(user preferences, 𝑇𝑜𝑝𝑃𝑜𝐼 )

return PoIs sorted by similarity-PoI

if in l,t, (w=rainy or w=stormy or w=snowy)

give priority to indoor places as sorted above

Step3:
else return previous results

Evaluation parameters: Precision, Recall and F-measure

3-1. Extracting preferences

Four main steps including preprocessing, semantic graph formation, clustering, and preferences extraction
are performed in this stage (see Fig. 1).
 Fig. 1- Extracting a user’s preferences

The pre-processing is performed to convert an initial document into a suitable processing form. Fig. 2 shows
the scheme of this step. The operations are described below.

Eliminating Extracting
PoS tagging Stemming
stop words nouns

Fig. 2- Pre-processing operations

Part of Speech (PoS) tagging: In this sub-stage, the constituents of the sentence, including noun, verb, etc.
are identified and tagged. This helps to extract information from sentences.
Stop words elimination: In a sentence, the words that do not have any specific meaning are called stop
words. By eliminating them, only the words with useful information remain.
Stemming: Stemming refers to the process of converting a word into its base or stem. In this sub-stage,
words such as cats and catlike are turned into their stem, i.e. cat. To this end, Wordnet is used, which keeps
the stem form of all words.
Extracting nouns: Nouns are the most informative constituents of a sentence. Using only nouns increases
the clustering efficiency in comparison with the use of all the words in a text(Fodeh, Punch, & Tan, 2011).
As a result, only the words that are tagged as noun are extracted in this sub-stage for further processes.

After pre-processing, the noun similarity matrix is constituted. It is a symmetric matrix in which rows and
columns correspond to the extracted nouns (as shown in Fig. 3). The semantic similarity between each pair
of entries forms the corresponding element in the matrix.

Noun 1 Noun 2 . . . Noun n

Noun 1

Noun 2
.
.
.

Noun n

Fig. 3- Nouns matrix

To constitute the matrix, a previously proposed hybrid semantic similarity measure (Wei, Lu, Chang, Zhou,
& Bao, 2015) is exploited. This measure considers all the direct and indirect relationships between concepts
in the Wordnet to increase the accuracy in calculating the similarity. It addresses the defects of the Wu-
Palmer semantic similarity (Wu & Palmer, 1994) (Lack of the direct relationships between concepts) and
the Extended gloss overlaps (Banerjee & Pedersen, 2003) (Lack of similarity between the concepts that
have a direct relationship in the structure of WordNet, but are not similar in their definitions).

Afterward, the elements of the matrix are normalized so that the method can be evaluated with different
semantic similarity measurement criteria. To this end, the values of all matrix elements are divided by the
maximum element.

Finally, the graph for this matrix is created, the vertices of which represent the extracted nouns of the
preprocessing stage and the weights of the edges represent the semantic similarity of the two meeting
vertices (nouns). If there is not semantic similarity between a pair of vertices, these two vertices are not
connected. Then, the edges that have a weight less than a specified threshold are eliminated. As a result,
the graph may be converted to several connected sub-graphs. Each sub-graph is considered as a cluster in
which each noun has a reasonable semantic similarity with at least one other noun.

Finally, each sentence is transferred to the clusters that contain any of its constituent words. Hence, based
on the nouns in a sentence, it might be placed in more than one cluster. For example, if the "X" cluster
contains a, b, c, and d nouns, and the "Y" cluster contains e and f, the “abg” sentence is located in the cluster
X, while the “afh” sentence is assigned to both clusters.
In continue, each cluster is scored based on the result of sentiment analysis as well as the frequency of
nouns. The main aim of sentiment analysis is to detect the feelings hidden in users’ comments and to
understand their thoughts regarding the subject (Mowlaei, Abadeh, & Keshavarz, 2020).
In this research, sentiment analysis is performed semantically with the help of the Sentiwordnet
(Baccianella, Esuli, & Sebastiani, 2010). Since a word in different situations may have different meanings
and sentimental loads, average of the positive as well as negative loads of synsets of the word is used as the
positive and negative load of that word, respectively. The score of each sentence is computed by deducting
the negative scores from the positive ones. The emotional load of emoticons is also taken into account.
Positive emoticons of each sentence are scored +1; while negative emoticons are scored -1 (Neidhardt et
al., 2017).
As equation (1) expresses, the sentiment analysis score of each cluster is equal to the average of sentiment
analysis score of its sentences.

∑𝑆𝑐𝑜𝑟𝑒 𝑜𝑓 𝑒𝑎𝑐ℎ 𝑠𝑒𝑛𝑡𝑒𝑛𝑐𝑒 𝑜𝑓 (𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖)

𝑆𝑐𝑜𝑟𝑒𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡 𝐴𝑛𝑎𝑙𝑦𝑠𝑖𝑠(𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖) = 𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑒𝑛𝑡𝑒𝑛𝑐𝑒𝑠 𝑖𝑛 (𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖)


In a text, repeated words (except for stop words) are usually more important than less frequent words
(Binwahlan, Salim, & Suanmali, 2010). Therefore, the frequency of nouns of each cluster has been involved
in scoring the clusters. Equation (2) shows the formula for computing the ith cluster score.

𝑆𝑐𝑜𝑟𝑒𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖 = 𝑇𝐹𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖 × 𝑆𝑐𝑜𝑟𝑒𝑆𝑒𝑛𝑡𝑖𝑚𝑒𝑛𝑡 𝐴𝑛𝑎𝑙𝑦𝑠𝑖𝑠(𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑖)

(2)

Where 𝑇𝐹 𝑐𝑙𝑢𝑠𝑡𝑒𝑟 𝑖 is equal to the total number of repetitions of nouns in the cluster i. For example, for a
cluster i with noun frequencies of 2, 2, 5, and 6, the value of 𝑇𝐹𝑐𝑙𝑢𝑠𝑡𝑒𝑟 𝑖 is 15.
Finally, the cluster that gains the highest score is considered as the selected cluster. In addition, sufficiently
closed clusters to the selected cluster (less than 10-point score difference) are also considered as selected
clusters. The set of nouns in these clusters represent the preferences of the user.

3-2. Extracting attractions features

After extracting users’ preferences, the features of tourism attractions should be extracted. In order to
consider the quality of recommendations, the attractions that have received less than 3 stars by tourists are
ignored. Besides, tourism attractions have different characteristics in different weather conditions. For
example, while the Alps in snowy days are known for snowy landscape and cold weather; in sunny days of
summer, they are known for lush landscape and moderate weather. Hence, for every tourism attraction, the
set of reviews of tourists, who have visited the attraction, are collected and preprocessed in five different
weather conditions including snowy, rainy, sunny, stormy and partly cloudy. Then, the top five most
repeated words of each attraction in each of these five weather conditions are extracted and are considered
as the attraction’s features. Fig. 4 shows the proposed steps.

Tourists
Location Extracting the top five Frequent words
reviews on
information nearby Pre- of the attraction
most repeated words
attractions processing
of the tourist attraction

Weather
information

Fig. 4- Extracting an attraction’s features

3-3. Recommendation system

The main idea of the proposed recommendation system is to compare a tourist’s preferences with the nearby
attractions’ features and return the most similar attractions. To compute the similarity of a user’s
preferences with an attraction’s features, the maximum similarity of any of the user’s preference elements
regarding all the features of the attraction is calculated. Therefore, these maximum similarities are averaged
for all preference elements of the user (Equation (3))(Haase, Siebes, & Van Harmelen, 2004).

1
𝑆𝑖𝑚(𝑃,𝐹) = |𝑃|∑𝑃 ∈𝑃
𝑀𝑎𝑥 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑃𝑖 , 𝐹𝑗) (3)
𝑖
𝐹𝑗 ∈ 𝐹
In the above Equation, P is the set of user’s preferences and F is the set of attraction’s features. The semantic
similarity criterion is the same as the measure used for the user preference extraction.

Finally, if the weather condition is snowy, rainy or stormy, indoor locations will be prioritized over outdoor
locations to reduce visiting troubles. On the other hand, if the weather is expected to be fine, the original
results will be shown. The general scheme of the recommendation is shown in Fig. 5.

User Weather
preferences information
Attractions’
Contextual features
information Visiting time User location

Wordnet Recommendation system Recommendations

Fig. 5- General scheme of attraction recommendation

4. Evaluation

The proposed recommender system has been developed by the python3 language and Anaconda1 platform.
As TripAdvisor is the most well-known tourism platform(Pantano, Priporas, & Stylos, 2017), which has
been used to collect a dataset for evaluating the proposed recommender system. The main part of the dataset
(training data) includes the reviews of 100 tourists with different ages and nationalities on various tourist
attractions in the course of six months (January-June 2018). The first travel of the users after this period
(from July 2018) along with all of its attraction visits are considered as the test data. Finally, the sentiment
analysis process has been performed using Sentiwordnet 3.0(Baccianella et al., 2010).At first, the proposed
sentiment analysis method is validated by comparing it with the approaches that are based on Support
Vector Machine (SVM) and Bayesian network (which are two popular methods in the field of sentiment

1 www.anaconda.org
analysis(Rana & Singh, 2016)). To this end, the reviews of tourists are categorized into either positive or
negative classes. Besides, this classification is also performed by considering ratings of users. In this case,
if the user has given a score of 4 or 5 bubbles (very good, excellent1), it is considered positive and if the
score is equal to 1 or 2 bubbles (terrible, poor), it is considered negative (Valdivia, Luzón, & Herrera,
2017).Afterwards, the percentage of positive and negative comments are computed by each approach and
the result is shown in Fig. 6. It reveals that the proposed approach achieves the most similar results to the
ratings provided by users.

Sentiment Analysis Score

100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
positive negative

User rating Proposed approach SVM Bayesian

Fig. 6- Validating the proposed sentiment analysis method

Next, the utilized hybrid semantic similarity measure, which is a combination of the two criteria of Wu-
Palmer (Wu & Palmer, 1994) and Extended Gloss Overlaps (Banerjee & Pedersen, 2003), is evaluated. To
this end, it is compared with each of these two methods in clustering the nouns of user comments. On the
other hand, the Cosine, Jaccard (Liu, Hu, Mian, Tian, & Zhu, 2014) and hybrid (Wei et al., 2015) semantic
similarity measures are investigated to calculate the similarity of user preferences with features of tourist
attractions.

φ φ
If the radius of user’s current city is φ km, three values of 4 , 2 and φ are considered as the allowed distance

for suggestion. Besides, three different modes are considered for evaluation, which are suggesting one
(Top1), three (Top3), and five (Top5) recommendations, respectively. For each of these scenarios, three

1
criteria of precision, recall, and f-measure are evaluated. Precision estimates what percentage of
recommendations are actually visited by the user. It should be noted for top3 and top5 that if the user visits
any of the recommended attractions, the recommendation is regarded as successful. Similarly, recall
indicates what percentage of actually visited attractions are recommended by the system. Figure 7 shows
the evaluation results for top1. The values of the horizontal axis (A-B) represent the semantic similarity
measures used for clustering nouns of user comments and the measure for similarity estimation of
preferences with features of attractions, respectively. In fact, A represents the measure used in the clustering
of nouns, and B represents the measure used in estimating the similarity between users preferences with
attractions features.

Precision of Top1 recommendation Reall of Top1 recommendation

φ/4 φ/2 φ φ/4 φ/2 φ
0.8 0.12
0.7 0.1
0.6 0.08
0.5 0.06
0.4 0.04
0.3 0.02
0.2 0
rd brid sine ard brid sine ard brid sine
ca c c
rd

id rid

O* ne

d
P* sine

W card

W brid

e
J- a -Hy -Co -Ja -Hy -Co -Ja -Hy -Co

sin
EG car

EG bri
ca

EG osi
b

id rid rid O* GO GO P** P P

Hy Hy

Hy

Hy
Co

Co
-Ja

-Ja

Ja
r
-C

b b b EG W W

*-
E E W
-

O-

O-

P-

P-
Hy Hy Hy
id

id
br

br

br

W
Hy

Hy

* EGO = Extended Gloss Overlaps * EGO = Extended Gloss Overlaps

** WP = Wu-Palmer ** WP = Wu-Palmer

F-measure of Top1 recommendation

φ/4 φ/2 φ
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
d

rid

rd

EG brid

rd

id

e
in

sin

sin
r

br
ca

ca

ca
yb

os

Hy

Hy
Co

Co
-Ja

-Ja

Ja
-H

-C

*-
O-

O-

P-

P-
id

O*
id

id

P*

W
br

EG
br

br

EG

W
Hy

Hy

Hy

* EGO = Extended Gloss Overlaps

** WP = Wu-Palmer

Fig7: Top1 recommendation (precision, recall, f-measure)

The results indicate the superiority of the hybrid criteria in comparison with other similarity metrics for
top1. In fact, the use of hybrid methods results in the best precision, recall, and f-measure for different radii.
Correspondingly, figures 8 and 9 show similar results for top3 and top5, respectively.

Precision of Top3 recommendation Recall of Top3 recommendation

φ/4 φ/2 φ φ/4 φ/2 φ
0.9 0.4
0.85 0.34
0.8
0.75 0.28
0.7 0.22
0.65 0.16
0.6
0.55 0.1
0.5 0.04
d

id

O* ine

O- id
P* sine

W card

W brid

id

O* ine

O- rid

*- e
W card

W brid

e
sin
br car

EG car

in

sin
br car

EG car
br

EG br

br

b
s

W Cos
Hy -Hy

Hy

Hy
o

Co

Co

Hy -Hy

Hy

Hy
o

Co
-Ja

-Ja

Ja

-Ja

-Ja

Ja
-C

-C
*-
O-

P-

P-

O-

P-

P-
id

id
id

id

id

id

P*
br

br
br

EG
br
EG

EG
W
Hy

Hy
Hy

* EGO = Extended Gloss Overlaps Hy

* EGO = Extended Gloss Overlaps
** WP = Wu-Palmer ** WP = Wu-Palmer

F-measure of top 3 recommendation

φ/4 φ/2 φ
0.6
0.5
0.4
0.3
0.2
0.1
0
d

rid

EG brid

rd

id

e
in

sin

sin
r

ar

br
ca

ca
yb

os

c
Hy

Hy
Co

Co
-Ja

-Ja

Ja
-H

-C

*-
O-

O-

P-

P-
id

O*
id

id

P*

W
br

EG
br

br

EG

W
Hy

Hy

Hy

* EGO = Extended Gloss Overlaps

** WP = Wu-Palmer

Fig. 8- Top3 recommendation (precision, recall, f-measure)

 Precision of Top5 recommendation Recall of Top5 recommendation
φ/4 φ/2 φ φ/4 φ/2 φ
0.95 1
0.9
0.85 0.8
0.8 0.6
0.75 0.4
0.7
0.65 0.2
0.6 0
rd

id rid

O* ine

O- rd

O- id

*- e
W card

W brid

id rid

O* ne

O- id

*- e
W card

W brid

e
in

sin

in

sin
ar

EG car
EG ybr
ca

EG aca

EG br
b

EG osi
EG os

W C os

W Cos
Hy -Jac
Hy -Hy

Hy

Co
-Ja

Ja

Hy -Hy

Hy

Hy

Co
-Ja

Ja
H
-C

-J

-C
P-

P-

O-

P-

P-
rid

id

id

id
P*

P*
br

br

br

br

br
b
Hy

Hy

Hy
* EGO = Extended Gloss Overlaps * EGO = Extended Gloss Overlaps
** WP = Wu-Palmer ** WP = Wu-Palmer

F-measure of Top5 recommendation

φ/4 φ/2 φ
0.9
0.86
0.82
0.78
0.74
0.7
d

id rid

O* ne

O- rd

O- id

*- e
W card

W brid

e
in

sin
br car

EG ybr
EG aca
EG osi
b

W Cos
Hy -Hy

Hy

Co
-Ja

Ja
H
-C

-J

P-

P-
id

id

P*
br

br
Hy

Hy

* EGO = Extended Gloss Overlaps

** WP = Wu-Palmer

Fig. 9- Top5 recommendation (precision, recall, f-measure)

When the number of recommendations increases, the precision (the possibility that the user visits at least
one of them) as well as recall (the possibility that a visited attraction has been recommended) increases.
Since f-measure is dependent on precision and recall, it is also greater for top5 comparing with top3, which
in turn yields better f-measure comparing with top1.

Finally, the proposed system is compared with two other similar systems proposed by Leal et al. (Leal et
al., 2017) and Loh et al. (Loh et al., 2003). Due to the similarity of the basis of these studies with our idea,
they have been used to evaluate the effectiveness of the proposed method. As mentioned in the related work
section, Loh et al. use the keywords of the user's reviews in a private chat and consider them as the user's
preferences and offer suggestions based on them. In the Leal's proposed method, user preferences and
features of tourist attractions are extracted using the topic modeling and suggestions are presented based on
the similarity between preferences and features.
In this research, f-measure is the most comprehensive criterion, hence the comparison is based on it. Fig.
10 shows the results of comparing f-measure for top1, top3, and top5 as well as the radius of φ/4. This
comparison is based on the highest f-measure score obtained in the evaluation.

According to the evaluation results, the proposed system has achieved desired values for the evaluation
measures. Also, since most recommender systems offer more than one recommendation, if the user receives
more offers, the proposed system will perform much better than similar systems.

F-measure
1

0.8

0.6

0.4

0.2

0
Top1 Top3 Top5

Loh, et al. Leal, et al. Proposed Method

Fig. 10- Comparison of the proposed system with similar methods

5. Conclusion

In this paper, a context-aware tourism recommendation system has been presented, which extracts a user’s
preferences by performing semantic clustering as well as sentiment analysis on their comments and reviews.
By sentiment clustering, the frequent concepts available in the user’s reviews are identified. Sentiment
analysis refines this list by distinguishing between preferences and unpopular items. Similarly, the features
of an attraction are extracted from aggregated reviews of users about it. Subsequently, nearby attractions
are ranked according to their similarity with the user’s preferences as well as current contextual information.
Finally, an experiment has been conducted on TripAdvisor website and reviews published by one hundred
users were extracted. For each user, the system has been trained by the data from the first 6 months, while
the first trip after this period along with all of its attractions visits were considered for testing the system.
Evaluation results have shown that the proposed system has outperformed comparable systems in terms of
f-measure, because it thoroughly leverages the sentiment analysis as well as context awareness potentials.
By analyzing sentiments, the proposed system is able to filter the frequent words that are disliked by the
user from their preferences list. As a result, the precision of the system increases comparing with previous
similar studies that have not investigated sentiments of comments. Moreover, by considering several
contextual elements, the proposed system has been able to recommend attractions that are adapted to the
current situation. As a result, the precision further increases comparing with previous works.

Although time, location, weather, users’ preferences, and attractions’ features have been exploited as vital
contextual information, other contextual information such as the traffic of routes, various conditions of
individuals as well as group users and their environment have been neglected. Augmenting the proposed
system by taking into account all the contextual information could enhance the recommendation process as
well as user satisfaction and convenience. As users usually travel and visit attractions in groups, extending
the proposed recommendation system for group scenarios is an important future research direction.
Acknowledgement
It is a pleasure for authors to thank Mr. Saeed hosseinabadi for his valuable remarks and helps.

References
Abbasi-Moud, Z., Vahdat-Nejad, H., & Mansoor, W. (2019). Detecting Tourist's Preferences by Sentiment Analysis
in Smart Cities. Paper presented at the 2019 IEEE Global Conference on Internet of Things (GCIoT).
Abel, F., Herder, E., Houben, G.-J., Henze, N & ,.Krause, D. (2013). Cross-system user modeling and personalization
on the social web. User Modeling and User-Adapted Interaction, 23(2-3), 169-209.
Abowd, G. D., Dey, A. K., Brown, P. J., Davies, N., Smith, M., & Steggles, P. (1999). Towards a better understanding
of context and context-awareness. Paper presented at the International symposium on handheld and
ubiquitous computing.
Alaei, A. R., Becken, S., & Stantic, B. (2019). Sentiment analysis in tourism: capitalizing on big data. Journal of
Travel Research, 58(2), 175-191.
Alrasheed, H., Alzeer, A., Alhowimel, A., & Althyabi, A. (2020). A Multi-Level Tourism Destination Recommender
System. Procedia Computer Science, 170, 333-340.
Anacleto, R., Figueiredo, L., Almeida, A., & Novais, P. (2014). Mobile application to provide personalized sightseeing
tours. Journal of Network and Computer Applications, 41, 56-64.
Baccianella, S., Esuli, A., & Sebastiani, F. (2010). Sentiwordnet 3.0: an enhanced lexical resource for sentiment
analysis and opinion mining. Paper presented at the Lrec.
Banerjee, S., & Pedersen, T. (2003). Extended gloss overlaps as a measure of semantic relatedness. Paper presented
at the Ijcai.
Bao, Y., Fang, H., & Zhang, J. (2014). Topicmf: Simultaneously exploiting ratings and reviews for recommendation.
Paper presented at the Twenty-Eighth AAAI conference on artificial intelligence.
Barranco, M. J., Noguera, J. M., Castro, J., & Martínez, L. (2012). A context-aware mobile recommender system
based on location and trajectory. In Management intelligent systems (pp. 153-162): Springer.
Binwahlan, M. S., Salim, N., & Suanmali, L. (2010). Fuzzy swarm diversity hybrid model for text summarization.
Information processing & management, 46(5), 571-588.
Borràs, J., Moreno, A., & Valls, A. ( .(2014Intelligent tourism recommender systems: A survey. Expert Systems with
Applications, 41(16), 7370-7389.
Cai, G., Lee, K., & Lee, I. (2018). Itinerary recommender system with semantic trajectory pattern mining from geo-
tagged photos. Expert Systems with Applications, 94, 32-40.
Chen, L., Zhang, L., Cao, S., Wu, Z., & Cao, J. (2020). Personalized itinerary recommendation: Deep and collaborative
learning with textual information. Expert Systems with Applications, 144, 113070.
Colomo-Palacios, R., García-Peñalvo, F. J., Stantchev, V., & Misra, S. (2017). Towards a social and context-aware
mobile recommendation system for tourism. Pervasive and Mobile Computing, 38, 505-515.
Esmaeili, L., Mardani, S., Golpayegani, S. A. H., & Madar, Z. Z. (2020). A novel tourism recommender system in the
context of social commerce. Expert Systems with Applications, 149, 113301.
Fodeh, S., Punch, B., & Tan, P.-N. (2011). On ontology-driven document clustering using core semantic features.
Knowledge and information systems, 28(2), 395-421.
Ghane’i-Ostad, M., Vahdat-Nejad, H., & Abdolrazzagh-Nezhad, M. (2018). Detecting overlapping communities in
LBSNs by fuzzy subtractive clustering. Social Network Analysis and Mining, 8(1), 23.
Haase, P., Siebes, R., & Van Harmelen, F. (2004). Peer selection in peer-to-peer networks with semantic topologies.
Paper presented at the International Conference on Semantics for the Networked World.
Hayashi, T., & Yoshida, T. (2018). Development of a Tour Recommendation System Using Online Customer Reviews.
Paper presented at the International Conference on Management Science and Engineering Management.
Kulkarni, S., & Rodd, S. F. (2020). Context Aware Recommendation Systems: A review of the state of the art
techniques. Computer Science Review, 37, 100255.
Leal, F., González–Vélez, H., Malheiro, B., & Burguillo, J. C. (2017). Semantic profiling and destination
recommendation based on crowd-sourced tourist reviews. Paper presented at the International Symposium
on Distributed Computing and Artificial Intelligence.
Liu, H., Hu, Z., Mian, A., Tian, H., & Zhu, X. (2014). A new user similarity model to improve the accuracy of
collaborative filtering. Knowledge-Based Systems, 56, 156-166.
Loh, S., Lorenzi, F., Saldaña, R., & Licthnow, D. (20 .(03A tourism recommender system based on collaboration and
text analysis. Information Technology & Tourism, 6(3), 157-165.
Lu, J., Wu, D., Mao, M., Wang, W., & Zhang, G. (2015). Recommender system application developments: a survey.
Decision Support Systems, 74, 12-32.
Lyu, D., Chen, L., Xu, Z., & Yu, S. (2020). Weighted multi-information constrained matrix factorization for
personalized travel location recommendation based on geo-tagged photos. Applied Intelligence, 50(3), 924-
938.
Majid, A., Chen, L., Chen, G., Mirza, H. T., Hussain, I., & Woodward, J. (2013). A context-aware personalized travel
recommendation system based on geotagged social media data mining. International Journal of
Geographical Information Science, 27(4), 662-684.
Missaoui, S. ,Kassem, F., Viviani, M., Agostini, A., Faiz, R., & Pasi, G. (2019). LOOKER: a mobile, personalized
recommender system in the tourism domain based on social media user-generated content. Personal and
Ubiquitous Computing, 23(2), 181-197.
Mowlaei, M. E., Abadeh, M. S., & Keshavarz, H. (2020). Aspect-based sentiment analysis using adaptive aspect-
based lexicons. Expert Systems with Applications, 148, 113234.
Neidhardt, J., Rümmele, N., & Werthner, H. (2017). Predicting happiness: user interactions and sentiment analysis in
an online travel forum. Information Technology & Tourism, 17(1), 101-119.
Pantano, E., Priporas, C.-V., & Stylos, N. (2017). ‘You will like it!’using open data to predict tourists' response to a
tourist attraction. Tourism Management, .438-430 ,60
Pantano, E., Priporas, C.-V., Stylos, N., & Dennis, C. (2019). Facilitating tourists' decision making through open data
analyses: A novel recommender system. Tourism Management Perspectives, 31, 323-331.
Pu, Z., Du, H., Yu, S., & Feng, D. .(2020)Improved Tourism Recommendation System. Paper presented at the
Proceedings of the 2020 12th International Conference on Machine Learning and Computing.
Rana, S., & Singh, A. (2016). Comparative analysis of sentiment orientation using SVM and Naive Bayes techniques.
Paper presented at the 2016 2nd International Conference on Next Generation Computing Technologies
(NGCT).
Renjith, S., Sreekumar, A., & Jathavedan, M. (2020). An extensive study on the evolution of context-aware
personalized travel recommender systems. Information processing & management, 57(1), 102078.
Smallwood, C. B., Beckley, L. E., & Moore, S. A. (2012). An analysis of visitor movement patterns using travel
networks in a large marine park, north-western Australia. Tourism Management, 33(3), 517-528.
Tumas, G., & Ricci, F. (2009). Personalized mobile city transport advisory system. Information and communication
technologies in tourism 2009, 173-183.
Vahdat-Nejad, H. (2014). Context-aware middleware: A review. In Context in computing (pp. 83-96): Springer.
Valdivia, A., Luzón, M. V., & Herrera, F. (2017). Sentiment analysis in tripadvisor. IEEE Intelligent Systems, 32(4),
72-77.
Wan, L., Hong, Y., Huang, Z., Peng, X., & Li, R. (2018). A hybrid ensemble learning method for tourist route
recommendations based on geo-tagged social networks. International Journal of Geographical Information
Science, 32(11), 2225-2246.
Wei, T., Lu, Y., Chang, H., Zhou, Q., & Bao, X. (2015). A semantic approach for text clustering using WordNet and
lexical chains. Expert Systems with Applications, 42(4), 2264-2275.
Wu, Z., & Palmer, M. (1994). Verbs semantics and lexical selection. Paper presented at the Proceedings of the 32nd
annual meeting on Association for Computational Linguistics.
Xiang, Z., Du, Q., Ma, Y., & Fan, W. (2017). A comparative analysis of major online review platforms: Implications
for social media analytics in hospitality and tourism. Tourism Management, 58, 51-65.
Yeh, D.-Y., & Cheng, C.-H. (2015). Recommendation system for popular tourist attractions in Taiwan using Delphi
panel and repertory grid techniques. Tourism Management, 46, 164-176.
Yochum, P., Chang, L., Gu, T., & Zhu, M. (2020). Linked Open Data in Location-Based Recommendation System
on Tourism Domain: A Survey. IEEE Access, 8, 16409-16439.
Yochum, P., Chang, L., Gu, T., Zhu, M., & Zhang, W. (2018). Tourist Attraction Recommendation Based on
Knowledge Graph. Paper presented at the International Conference on Intelligent Information Processing.
Zheng, X., Ding, W. ,Lin, Z., & Chen, C. (2016). Topic tensor factorization for recommender system. Information
Sciences, 372, 276-293.
Zheng, X., Luo, Y., Sun, L., Zhang, J., & Chen, F. (2018). A tourism destination recommender system using users’
sentiment and temporal dynamics. Journal of Intelligent Information Systems, 51(3), 557-578.
 Highlights

 Users reviews on tourism networks are processed to extract their preferences.

 Attractions aggregated reviews are processed to extract their features.
 A personalized tourism recommendation system is proposed.
 The proposed recommendation system is context-aware.
Zahra Abbasi-Moud: Conceptualization, Methodology, Software, Original draft
preparation
Hamed Vahdat-Nejad: -Supervision- Conceptualization, Methodology- Reviewing
and Editing,
Javad Sadri: Supervision- Methodology