A Lexicon-based Collaborative Filtering Approach for Recommendation

Systems

Mara Deac-Petrus¸el

Faculty of Mathematics and Computer Science, Babes¸-Bolyai University, Cluj-Napoca, Romania

Keywords:

Recommendation Systems, Collaborative Filtering, K Nearest Neighbours, Vader Lexicon, Text-based

Reviews, Sentiment Score.

Abstract:

Users purchasing items from e-commerce websites are expressing their satisfaction and sentiment about their

acquisition using text-based reviews and numerical ratings. Traditional collaborative ﬁltering techniques are

entirely dependent on the users’ scalar ratings, which are lacking any semantic explanation of the users’

preferences. This approach was designed to explore the text-based item evaluation using a Sentiment Anal-

ysis Lexicon. The proposed lexicon-based k nearest neighbors collaborative ﬁltering technique replaces the

numerical rating with a computed sentiment rating in the neighborhood determination step. The conducted

experiments reveal that the resulting text-based recommendation system produces reliable values in terms of

mean absolute error and root mean square error and accurate recommendations for users.

1 INTRODUCTION

E-commerce websites are important sources of data

generation. People are buying items based on the

reviews of other peers. Users evaluate the pur-

chased items, expressing their satisfaction and opin-

ions through numerical ratings and text-based re-

views. Classic recommendation systems produce sug-

gestions for users using the given numerical ratings

for items. Scalar ratings are not able to offer a seman-

tic explanation about the users’ preferences. There-

fore, recommendation systems are facing the chal-

lenge of efﬁciently analyzing information in textual

form.

The proposed approach was designed to capture

the users’ interests from the text-based items’ reviews

to produce good rating predictions for items and ac-

curate generated recommendations for users. The

items’ descriptions are passed to a Sentiment Anal-

ysis Lexicon, which outputs a sentiment score indi-

cating the polarity of the text (positive, negative, or

neutral). Based on the sentiment score, a k-nearest

neighbor (kNN) user-based collaborative ﬁltering al-

gorithm was applied. The recommendation system

uses solely the sentiment scores (called sentiment rat-

ings), instead of the numerical ratings. Results have

proven a positive impact of the text-based approach

on the performance of the recommendation system.

The original contributions of the papers are:

• Analysing the textual information of an item and

applying the Vader Sentiment Analysis Lexicon

(Hutto and Gilbert, 2014) to obtain the sentiment

rating;

• Integrate the lexicon-based data into the k nearest

neighbors user-based collaborative ﬁltering algo-

rithm;

The rest of this paper is organized as follows. Chap-

ter two surveys the state-of-art approaches that in-

tegrate the rich information contained in text-based

items’ descriptions in the recommendation process.

The third chapter describes a detailed methodology

used to design this approach. The fourth chapter

presents the experimental setups conducted and gives

an overview of the selected data sets in terms of

their features and volume. Also, the evaluation met-

rics used in the textual recommendation system are

briefed and comparisons to other approaches are dis-

cussed. In the last chapter, the overall summary of

this work is presented, highlighting the main ﬁndings,

results, and future work plans.

2 RELATED WORK

This section offers an overview of several state-of-

the-art approaches where text-based reviews are ex-

plored and integrated into the collaborative ﬁltering

Deac-Petru¸sel, M.

A Lexicon-based Collaborative Filtering Approach for Recommendation Systems.

DOI: 10.5220/0010801200003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 203-210

ISBN: 978-989-758-547-0; ISSN: 2184-433X

203

recommendation process to improve rating predic-

tion.

(Terzi et al., 2014) proposes a text-based user-

kNN algorithm that uses text-based reviews instead

of numerical ratings to compute the users’ similar-

ity. The idea is to determine the similarity between

two users by computing the similarity between re-

views’ words for every item reviewed by both users.

The text-based user-kNN is compared with several

approaches using numerical ratings in the rating pre-

diction step. For the numerical experiments, two data

sets are used: RottenTomatoes and an AudioCD from

AmazonProductReviews. Slightly better results are

obtained for root mean square error (RMSE) between

the actual and the predicted ratings in the text-based

user-kNN approach over the ratings-based ones.

(Poirier et al., 2010) determines sentiment scores

from text-based reviews using a Na

ıve Bayes model.

As a ﬁrst step, the text-based reviews are analyzed

and text mining techniques are applied in order to

build the user-item-rating matrix. Reviews are clas-

siﬁed into two sentiment classes: positive and neg-

ative using the KHIOPS tool (Boull

e, 2007). Then,

the item-based collaborating ﬁltering algorithm is ap-

plied to generate the recommendations. Experiments

are conducted using Flixter, Netﬂix, and IMDB data

sets. RMSE is used as an evaluation measure.

(Ma et al., 2017) designed an original user-

preference-based collaborative ﬁltering (UPCF) ap-

proach to exploit free-text online reviews to retrieve

users’ preferences. Firstly, aspect-level opinions min-

ing techniques were applied to transform the free-text

reviews into structured aspect opinions. Next, the user

preferences were determined on one hand from the

aspect importance and, on the other hand, from the

aspect need. The aspect importance means that opin-

ions on important aspects are more inﬂuential to the

overall ratings than other aspects, and uses the sim-

ilarity between the opinions on one aspect and the

overall ratings. The aspect need is calculated as the

difference between the opinions of a user on an as-

pect and those of other users, which indicates the dif-

ferentiated needing level on this aspect with respect

to the user. Based on this, a user-based collaborative

ﬁltering approach is designed so that the users’ aspect

preferences are integrated to calculate the similarities

between users.

(Musto et al., 2017) implemented a user and item-

based collaborative ﬁltering approach that includes

aspect opinion data. For both user and item-based

use cases, aspect-based user/item distances are cal-

culated using the sentiment ratings extracted from the

reviews’ aspects. The similarity between users/items

is determined based on the inverse of the users/items

distances and ratings’ predictions are computed using

the collaborative ﬁltering algorithm.

3 SYSTEM ARCHITECTURE

As exempliﬁed in Chapter 2, the text-based items’

descriptions reveal more valuable information com-

pared to the plain numerical ratings for the recom-

mendation process. The focus of the proposed ap-

proach is to make use solely of the textual information

when building the recommendation system, regard-

less of the numerical ratings. The textual input is ex-

ploited using a lexicon-based technique to determine

the polarity score of a review. The resulted scores

are the sentiment ratings taken into consideration for

the user-based kNN collaborative ﬁltering algorithm.

After the data collection phase, the text-based items’

reviews serve as input for a sentiment lexicon that de-

termines a sentiment rating for an item. The data set

enhanced with the computed sentiment rating is fur-

ther passed to a recommendation system.

3.1 Data Pre-processing

The proposed recommendation system handles tex-

tual information, therefore, a data cleansing process

was applied to the input data sets before being used

by the sentiment lexicon. The following techniques

have been applied:

• Removal of punctuation and stop words;

• Lower-casing;

• Removal of URLs;

• Stemming

3.2 Sentiment Lexicon

The proposed approach uses, for the sentiment anal-

ysis task, a sentiment lexicon, which was selected

based on the complex and thorough comparison pre-

sented in (Hutto and Gilbert, 2014). The Vader Sen-

timent Lexicon was compared to several ones from

literature (Linguistic Inquiry Word Count, General

Inquirer, Affective Norms for English Words, Sen-

tiWordNet, SenticNet, Word-Sense Disambiguation)

and produced,in most cases, the best results.

Vader (Valence Aware Dictionary and Sentiment

Reasoner) lexicon (Hutto and Gilbert, 2014) is a rule-

based sentiment analysis tool based on a dictionary

that maps words to positive, neutral, or negative sen-

timent scores. The sum of all these scores deﬁnes a

compound score which is normalized between -1 and

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

204

1. A value closer to -1 indicates a negative sentiment

for the item’s review, while a value closer to 1 a pos-

itive one. The sentiment score of a review is deter-

mined as follows:

score(review) =

∑

i=1

score(word

), (1)

where

• n is the number of words of a review.

• score(word

) is the sentiment score of the i

word, based on the Vader lexicon.

3.3 Recommendation Process

The data set containing in addition the reviews with

sentiment scores represents the input data for the rec-

ommendation system. The kNN collaborative ﬁl-

tering algorithm is then applied as recommendation

technique, as follows (Petrusel and Limboi, 2019):

• The most similar k users (called neighbors) for

a target user are determined based on the calcu-

lated users’ similarity. Several similarity mea-

sures from literature can be used at this step.

• For not yet reviewed items of the target user, the

rating prediction is computed and top-n recom-

mendations are generated.

To determine the unknown rating prediction for an

item i and the target user a, the following formula is

used (Victor et al., 2011):

a,i

= ¯score

∑

u∈R

a,u

(score

u,i

− ¯score

)

∑

u∈R

a,u

(2)

where:

• ¯score

is the mean lexicon-based score rating

given by target user a for other items than i.

• ¯score

is the mean lexicon-based score given by

user u over all items.

• w

a,u

is the similarity between the two users.

• score

u,i

is the lexicon-based score determined for

item i and user u.

• R

is the set of users that rated item i positively.

3.4 Evaluation Process & Conﬁdence

Intervals

The recommendation process is evaluated by comput-

ing the Mean Absolute Error (MAE) and Root Mean

Square Error (RMSE) (Isinkaye et al., 2015) for a user

u, based on the following formulas:

MAE(u) =

∑

i=1

u,i

− score

u,i

(3)

and

RMSE(u) =

∑

i=1

u,i

− score

u,i

(4)

where p

u,i

and score

u,i

are the predicted score, re-

spectively the actual score (determined via lexicon)

of user u for item i and N

is the number of recom-

mended items.

Moreover, 95% conﬁdence intervals are deter-

mined to ﬁnd out the value ranges for the model per-

formance. The conﬁdence interval is a statistic in-

terval for measuring the uncertainty on an estimate.

It measures the range of values, from the given data,

that includes the true values for estimating good sug-

gestions. A smaller conﬁdence interval means a more

precise estimation in comparison with a large one.

4 EXPERIMENTAL SETUP

To highlight the value-added by the proposed lexicon-

based K Nearest Neighbors collaborative ﬁltering

approach in improving the rating prediction accu-

racy, several numerical experiments were conducted

on three data sets containing text-based reviews for

items.

For the neighborhood determination in the kNN,

various popular similarity measures from litera-

ture were applied: Pearson Correlation Coefﬁcient

(PCC), Cosine (COS), Euclidean (EUC), Constrained

Pearson Coefﬁcient (CPC), Spearman Rank Coefﬁ-

cient (SRC), Jaccard Similarity (JAC) (Agarwal and

Chauhan, 2017), (Sondur et al., 2016) and PIP (Ahn,

2008). Independent scenarios are designed for differ-

ent values of k (the neighbourhood size) and n (the

number of generated recommendations).

In the evaluation process, the MAE and RMSE

measures are computed to establish the accuracy of

the generated recommendations.

4.1 Data Sets

For the numerical experiments, three data sets are

used: Amazon Fashion (Yan et al., 2019), Rotten

Tomato Critic Reviews (Firmanto et al., 2018) and

Dataﬁniti Product Reviews (Zahid-samza595, 2020).

The Amazon Fashion data set contains 100 000 re-

views and several features, such as: review time, re-

viewer name, review text.

A Lexicon-based Collaborative Filtering Approach for Recommendation Systems

205

The Rotten Tomato Critic reviews has 50 000

movies & TV reviews. It is composed of features like

publisher name, review type, review content or review

date.

The Dataﬁniti Product Reviews has 3000 wine,

beer and liquor reviews described by the business

name, brand, category, review text, review date or

username.

4.2 Amazon Fashion Data Set Results

In tables 1, 2, 3, 4, 5 and 6 the results obtained for

several scenarios are presented. The best values for

MAE and RMSE for the lexicon-based collaborative

ﬁltering are achieved using PIP similarity measure for

the neighborhood size k = 10 and number of recom-

mendations n = 3.

Moreover, the 95% conﬁdence intervals were

calculated and the best results were obtained for the

scenario considering the neighborhood size k = 10

and number of recommendations n = 3 and are pre-

sented in table 7. Results show that MAE was at

x + / − y for 95% conﬁdence interval, where x is the

lower bound of the interval and y is the upper bound.

4.3 Rotten Tomato Critic Reviews Data

Set Results

For the Rotten Tomato Critic Reviews data set (Fir-

manto et al., 2018), the results for all scenarios setups

are presented in tables 8, 9, 10, 11, 12 and 13. Best

values were achieved when using the PIP similarity

measure with neighbourhood size k = 10 and number

of generated recommendations n = 3.

Table 14 reveals the 95% conﬁdence interval com-

puted for the best scenario with neighbourhood size

k = 10 and number of generated recommendations

n = 3.

4.4 Dataﬁniti Product Reviews Data Set

Results

Tables 15, 16, 17, 18, 19 and 20 showcase the results

obtained for all the test scenarios, for the Dataﬁniti

Product Reviews data set. Best values were achieved

when using the Spearman Rank Coefﬁcient in the

lexicon-based kNN collaborative ﬁltering approach

for the neighbourhood size k = 5 and number of gen-

erated recommendations n = 3.

Table 21 presents the 95% conﬁdence interval for

the best scenario when using the neighbourhood size

k = 5 and number of generated recommendations n =

4.5 Comparisons and Discussions

This section offers an overview of the MAE values ob-

tained for the best scenarios for each data set. When

comparing the results presented in ﬁgures 2, 3 and 4,

it can be observed that the best performance in terms

of MAE was achieved for the Dataﬁniti Product Re-

views (Zahid-samza595, 2020) using the Spearman

Rank Coefﬁcient.

Moreover, the proposed lexicon-based approach is

compared to another text-based kNN collaborative ﬁl-

tering approach described in (Terzi et al., 2014), in

terms of Root Mean Square Error performance mea-

sure. Both approaches use text-based reviews instead

of numerical ones and the experiments are conducted

on the Rotten Tomato Critic Reviews data set. Al-

though both approaches make use of textual items’

descriptions, there is a difference in the sentiment

score deﬁnition (substituting the numerical rating).

(Terzi et al., 2014) computes the distance between two

words based on the shortest distance between them,

while in the proposed approach the sentiment score is

obtained based on the information derived from the

Vader Lexicon (Hutto and Gilbert, 2014).

Results are presented in table 22. Even though

the quantitative results in (Terzi et al., 2014) are bet-

ter, the presented approach is different from a quali-

tative point of view, using a lexicon-based collabora-

tive ﬁltering technique. The proposed technique has

value especially from the semantic point of view, con-

sidering words’ polarities (positive, negative, neutral)

compared to (Terzi et al., 2014), which is based on the

set of common words. Overall, this comparison high-

lights the fact that the presented approach generates

good and trustworthy results and conﬁrms again that

text-based reviews indeed offer valuable information

for the recommendation process.

5 TABLES & FIGURES

Table 1: Amazon Fashion k=5 and n=3.

Similarity Measure MAE RMSE

PCC 0.45 0.67

COS 1.19 1.07

EUC 1.90 1.70

CPC 0.75 0.85

SRC 0.20 0.31

JAC 0.29 0.49

PIP 0.10 0.30

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

206

Table 2: Amazon Fashion k=5 and n=5.

Similarity Measure MAE RMSE

PCC 0.41 0.64

COS 1.04 1.002

EUC 1.90 1.70

CPC 0.86 0.91

SRC 0.20 0.30

JAC 0.29 0.49

PIP 0.10 0.31

Table 3: Amazon Fashion k=5 and n=10.

Similarity Measure MAE RMSE

PCC 0.34 0.58

COS 1.03 0.99

EUC 1.80 1.70

CPC 0.72 0.84

SRC 0.20 0.30

JAC 0.29 0.49

PIP 0.10 0.31

Table 4: Amazon Fashion k=10 and n=3.

Similarity Measure MAE RMSE

PCC 0.45 0.67

COS 0.70 0.82

EUC 0.77 0.66

CPC 0.73 0.84

SRC 0.13 0.27

JAC 0.29 0.48

PIP 0.10 0.19

Table 5: Amazon Fashion k=10 and n=5.

Similarity Measure MAE RMSE

PCC 0.42 0.64

COS 0.86 0.90

EUC 1.77 1.66

CPC 0.85 0.90

SRC 0.20 0.32

JAC 0.29 0.49

PIP 0.10 0.29

6 CONCLUSIONS & FUTURE

WORK

Recommendation systems are important tools for

deﬁning suggestions for users. Lately, a lot of effort

was put into incorporating the text-based reviews in

the recommendation process. The scope was to en-

Table 6: Amazon Fashion k=10 and n=10.

Similarity Measure MAE RMSE

PCC 0.35 0.59

COS 0.83 0.89

EUC 1.77 1.66

CPC 0.72 0.84

SRC 0.20 0.32

JAC 0.29 0.49

PIP 0.10 0.29

Table 7: Amazon Fashion 95% CI.

Similarity Measure 95% CI

PCC (0.447,0.453)

COS (0.697,0.703)

EUC (0.767,0.773)

CPC (0.727,0.733)

SRC (0.128,0.132)

JAC (0.287,0.293)

PIP (0.098,0.102))

Table 8: Rotten Tomato Critic Reviews k=5 and n=3.

Similarity Measure MAE RMSE

PCC 0.78 0.77

COS 0.57 0.59

EUC 0.66 1.70

CPC 0.80 0.76

SRC 0.39 0.52

JAC 0.98 0.85

PIP 0.71 0.71

Table 9: Rotten Tomato Critic Reviews k=5 and n=5.

Similarity Measure MAE RMSE

PCC 0.85 0.82

COS 0.69 0.69

EUC 1.05 0.88

CPC 0.85 0.79

SRC 0.46 0.58

JAC 0.78 0.76

PIP 0.47 0.59

hance the classical collaborative ﬁltering algorithms

and to explore the power of these descriptions. The

text-based reviews describe the user’s opinions and

feelings about items more accurately than a numeri-

cal value and, therefore, the text-based techniques can

produce accurate items’ rating predictions and rec-

ommendations. With this in mind, a recommenda-

tion system that incorporates the textual information

A Lexicon-based Collaborative Filtering Approach for Recommendation Systems

207

Table 10: Rotten Tomato Critic Reviews k=5 and n=10.

Similarity Measure MAE RMSE

PCC 0.68 0.71

COS 0.64 0.66

EUC 0.70 0.72

CPC 0.84 0.79

SRC 0.54 0.62

JAC 0.49 0.59

PIP 0.39 0.53

Table 11: Rotten Tomato Critic Reviews k=10 and n=3.

Similarity Measure MAE RMSE

PCC 0.72 0.71

COS 0.44 0.55

EUC 0.54 0.61

CPC 0.49 0.58

SRC 0.57 0.65

JAC 0.97 0.83

PIP 0.20 0.32

Table 12: Rotten Tomato Critic Reviews k=10 and n=5.

Similarity Measure MAE RMSE

PCC 0.62 0.67

COS 0.62 0.63

EUC 0.66 0.70

CPC 0.59 0.65

SRC 0.53 0.62

JAC 0.53 0.62

PIP 0.24 0.42

Table 13: Rotten Tomato Critic Reviews k=10 and n=10.

Similarity Measure MAE RMSE

PCC 0.64 0.68

COS 0.64 0.67

EUC 0.61 0.67

CPC 0.81 0.76

SRC 0.51 0.61

JAC 0.67 0.70

PIP 0.55 0.64

about an item by a user was proposed. In comparison

with the majority of approaches presented in 2 that

use machine learning algorithms, the proposed one

focuses on a lexicon-based kNN collaborative ﬁlter-

ing technique. The text-based review is processed by

the Vader Lexicon (Hutto and Gilbert, 2014), which

computes the sentiment rating. Then, the data set aug-

mented with the sentiment rating is used as input for

Table 14: Rotten Tomato Critic Reviews 95% CI.

Similarity Measure 95% CI

PCC (0.717,0.723)

COS (0.437,0.443)

EUC (0.537,0.543)

CPC (0.487,0.493)

SRC (0.567,0.573)

JAC (0.969,0.971)

PIP (0.198,0.202))

Table 15: Dataﬁniti Product Reviews k=5 and n=3.

Similarity Measure MAE RMSE

PCC 0.71 0.74

COS 1.73 1.22

EUC 1.66 1.61

CPC 1.21 1.01

SRC 0.04 0.09

JAC 0.74 0.82

PIP 0.06 0.17

Table 16: Dataﬁniti Product Reviews k=5 and n=5.

Similarity Measure MAE RMSE

PCC 0.98 0.82

COS 1.60 1.19

EUC 1.73 1.63

CPC 1.04 1.94

SRC 0.10 0.22

JAC 0.74 0.82

PIP 0.06 0.16

Table 17: Dataﬁniti Product Reviews k=5 and n=10.

Similarity Measure MAE RMSE

PCC 0.89 0.81

COS 1.63 1.20

EUC 1.73 1.63

CPC 1.05 1.94

SRC 0.10 0.22

JAC 0.74 0.82

PIP 0.06 0.16

the kNN collaborative ﬁltering algorithm.

The results obtained in the conducted numerical

experiments show that the presented approach can be

successfully used to solve recommendation tasks, for

data sets containing text-based user reviews. As fu-

ture work, the approach could be extended to also

consider different types of review elements besides

words, such as review topics or aspect opinions.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

208

Table 18: Dataﬁniti Product Reviews k=10 and n=3.

Similarity Measure MAE RMSE

PCC 1.02 0.82

COS 1.57 1.76

EUC 1.40 1.52

CPC 0.88 0.87

SRC 0.08 0.21

JAC 0.58 0.70

PIP 0.09 0.24

Table 19: Dataﬁniti Product Reviews k=10 and n=5.

Similarity Measure MAE RMSE

PCC 1.01 0.82

COS 1.47 1.16

EUC 1.11 1.41

CPC 1.05 0.90

SRC 0.15 0.27

JAC 0.63 0.76

PIP 0.16 0.33

Table 20: Dataﬁniti Product Reviews k=10 and n=10.

Similarity Measure MAE RMSE

PCC 0.97 0.83

COS 1.69 1.23

EUC 1.12 1.42

CPC 1.05 0.94

SRC 0.15 0.27

JAC 0.63 0.75

PIP 0.16 0.34

Table 21: Dataﬁniti Product Reviews 95% CI.

Similarity Measure 95% CI

PCC (0.694,0.726)

COS (0.422,0.458)

EUC (0.522,0.558)

CPC (0.472,0.508)

SRC (0.552,0.588)

JAC (0.964,0.976)

PIP (0.186,0.214))

Table 22: Lexicon-based CF approach vs. (Terzi et al.,

2014) approach.

Approach RMSE

Lexicon-based CF 0.32

(Terzi et al., 2014) 0.14

Figure 1: MAE for Amazon Fashion, k=10 and n=3.

Figure 2: MAE for Rotten Tomato Critic Reviews, k=10

and n=3.

Figure 3: MAE for Dataﬁniti Product Reviews, k=5 and

n=3.

REFERENCES

Agarwal, A. and Chauhan, M. (2017). Similarity measures

used in recommender systems: a study. International

Journal of Engineering Technology Science and Re-

search IJETSR, ISSN, pages 2394–3386.

Ahn, H. J. (2008). A new similarity measure for collabo-

rative ﬁltering to alleviate the new user cold-starting

problem. Information sciences, 178(1):37–51.

A Lexicon-based Collaborative Filtering Approach for Recommendation Systems

209

Boull

e, M. (2007). Compression-based averaging of selec-

tive naive bayes classiﬁers. The Journal of Machine

Learning Research, 8:1659–1685.

Firmanto, A., Sarno, R., et al. (2018). Prediction of movie

sentiment based on reviews and score on rotten toma-

toes using sentiwordnet. In 2018 International Semi-

nar on Application for Technology of Information and

Communication, pages 202–206. IEEE.

Hutto, C. and Gilbert, E. (2014). Vader: A parsimonious

rule-based model for sentiment analysis of social me-

dia text. In Proceedings of the International AAAI

Conference on Web and Social Media, volume 8.

Isinkaye, F. O., Folajimi, Y., and Ojokoh, B. A. (2015). Rec-

ommendation systems: Principles, methods and eval-

uation. Egyptian informatics journal, 16(3):261–273.

Ma, Y., Chen, G., and Wei, Q. (2017). Finding users pref-

erences from large-scale online reviews for person-

alized recommendation. Electronic Commerce Re-

search, 17(1):3–29.

Musto, C., de Gemmis, M., Semeraro, G., and Lops, P.

(2017). A multi-criteria recommender system exploit-

ing aspect-based sentiment analysis of users’ reviews.

In Proceedings of the eleventh ACM conference on

recommender systems, pages 321–325.

Petrusel, M.-R. and Limboi, S.-G. (2019). A restaurants

recommendation system: Improving rating predic-

tions using sentiment analysis. In 2019 21st Inter-

national Symposium on Symbolic and Numeric Al-

gorithms for Scientiﬁc Computing (SYNASC), pages

190–197. IEEE.

Poirier, D., Fessant, F., and Tellier, I. (2010). Re-

ducing the cold-start problem in content recom-

mendation through opinion classiﬁcation. In 2010

IEEE/WIC/ACM International Conference on Web In-

telligence and Intelligent Agent Technology, volume 1,

pages 204–207. IEEE.

Sondur, M. S. D., Chigadani, M. A. P., and Nayak, S.

(2016). Similarity measures for recommender sys-

tems: a comparative study. Journal for Research, 2(3).

Terzi, M., Rowe, M., Ferrario, M.-A., and Whittle, J.

(2014). Text-based user-knn: Measuring user similar-

ity based on text reviews. In International Conference

on User Modeling, Adaptation, and Personalization,

pages 195–206. Springer.

Victor, P., Cornelis, C., De Cock, M., and Teredesai, A. M.

(2011). Trust-and distrust-based recommendations

for controversial reviews. IEEE Intelligent Systems,

(1):48–55.

Yan, C., Chen, Y., and Zhou, L. (2019). Differentiated fash-

ion recommendation using knowledge graph and data

augmentation. Ieee Access, 7:102239–102248.

Zahid-samza595, S. (2020). Sentiment analysis of hotel

reviews-performance evaluation of machine learning

algorithms.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

210