Analysis of economic satisfa=
ction using
machine learning models and explainable artificial intelligence
|
Luiz Fernando Menegazzo Ferrey=
ra |
=
Estudante de Engenharia de Produção. Universidade Tecnológica Feder=
al
do Paraná – Campus Londrina (UTFPR) – Brasil. luizferreyr=
a@alunos.utfpr.edu.br |
|
Yasser
Bulaty Tauil http=
s://orcid.org/0009-0001-4804-2416 |
Estudante de Engenharia de Produção.
Universidade Tecnológica Federal do Paraná – Campus Londrina (UTFPR) –
Brasil. yasser@alunos.utfpr.edu.br |
|
Helton
Messias Adigneri |
Bacharel em Engenharia de Produção.
Universidade Estadual de Maringá – Campus Maringá (UEM) – Brasil. =
pg405633@uem.br |
|
Bruno
Samways dos Santos |
Doutor em Engenharia de Produção e
Sistemas. Universidade Tecnológica Federal do Paraná – Campus Londrina
(UTFPR) – Brasil. =
brunosantos@utfpr.edu.br |
|
Rafael
Lima |
Doutor em Engenharia de Produção.
Universidade Tecnológica Federal do Paraná – Campus Londrina (UTFPR) –
Brasil. rafaelhlima@utfpr.edu.br |
ABSTRACT
The economic satisfaction of a nation can reflect citizens' percepti=
ons
of their government's performance, and machine learning models can help unc=
over
non-trivial information from such data. In this context, this article aimed=
to
analyze the satisfaction of Latin American citizens with their country's
economy. To achieve this, six traditional classifier algorithms and four
ensemble models were used, with a final application of an explainable method
(SHapley Additive exPlanations, SHAP) to analyze the key factors contributi=
ng
to economic satisfaction. The models were trained and tested on a dataset
comprising data from the 2020 and 2023 Latinobarómetro surveys, totaling 27=
,600
instances in the final set. As a result, it was found that the Random Forest
was the best individual model, while the stacking ensemble achieved the best
performance in classifying between “satisfied” and “dissatisfied” citizens.=
The
SHAP method revealed that “satisfaction with democracy” and “perception of =
the
country's progress” are the main factors influencing economic satisfaction.
This study offers insights for public managers on how to improve their
citizens' economic satisfaction.
Keywords: Economic satisfaction. Machine learning. =
Explainable
Artificial Intelligence. Latin America.
RESUMO
A satisfação econômica de uma nação pode refletir a percepção dos cidadãos sobre o desempenho de seus respectivos governos, e modelos de aprendizado de máquina podem auxiliar na descoberta = de informações não triviais contidas em dados desta natureza. Neste sentido, o objetivo deste artigo foi analisar a satisfação dos cidadãos latino-america= nos sobre a economia do seu país. Para isso, foram utilizados seis algoritmos classificadores tradicionais e mais quatro modelos ensemble, com a aplicação final de um método explicativo (SHapley Additive exPlanations<= /i>, SHAP), analisando os principais fatores que contribuem para a satisfação econômica. Os modelos foram treinados e testados em um conjunto de dados composto pelos anos de 2020 e 2023 da pesquisa do Latinobarómetro, totaliza= ndo 27.600 instâncias no conjunto final. Como resultado, verificou-se que a Floresta Aleatória foi o melhor modelo individual, enquanto o stacking ensemble obteve o melhor desempenho para a classificação entre “satisfeitos” e “insatisfeitos”. O método SHAP mostrou que a “satisfação co= m a democracia” e a “percepção sobre o progresso do país” sãos os principais fatores que influenciam na satisfação econômica. Este trabalho oferece cami= nhos nos quais gestores públicos podem atuar para a melhoria da satisfação econô= mica de seus cidadãos.
=
Palavras-chave:
Satisfação econômica. Aprendizado de máquina.
Inteligência Artificial Explicativa. América Latina.
Recebido em 31/08/2024. Apr= ovado em 04/11/2024. Avaliado pelo sistema double blind peer review. Publicado conforme normas da APA.
=
https://doi.org/10.22279/navus.v15.200=
6
1 INTRODUCTION
The advancement in understanding satisfaction involves recognizing t=
he
"object of satisfaction," as it is asserted that the satisfaction=
a
person feels with life as a whole is distinct from specific satisfaction wi=
th
work, marriage, or housing
The analysis of satisfaction has been researched since 1989, beginni=
ng
with a study in Sweden to measure consumer satisfaction using the Swedish
Customer Satisfaction Barometer (SCSB)
Regarding the general population of a nation, surveys on satisfaction
with services within a country are key in evaluating a government, revealing
judgments about the quality of services offered to citizens thus far. Howev=
er,
general satisfaction is multifactorial and subjective, making this task more
complex
ML algorithms are being widely utilized in various fields to analyze
satisfaction, including healthcare, products and services, economics, and e=
ducation,
among others. For example, authors such as
In Latin America, data from the Latinobarómetro survey evaluates pub=
lic
opinion in 18 countries concerning democracy, economy, and society. With th=
is
publicly available dataset, recent studies by
In this context, the present article seeks to analyze the applicatio=
n of
machine learning techniques to classify data concerning economic satisfacti=
on
in Latin American countries, including ensemble classifier models (bagging,
boosting, voting, stacking) and the SHAP technique (SHapley Additive exPlan=
ations)
to evaluate predictor variables. In addition to the use of algorithms, the
study also compares different years of questionnaire application in Latin
America, aiming to identify potential variations in economic satisfaction
across the continent.
Following this introductory section, the article comprises four main
sections. Section 2 outlines concepts related to data mining, machine learn=
ing,
ensemble classifiers, and interpretive models. The third section details the
research sequence, including the treatment of the datasets used and the
analysis of base algorithm hyperparameters. The fourth section presents the
results obtained by the classifiers, offering comparisons and interpretatio=
ns
of predictor variables through the SHAP model. Finally, the conclusion and
suggestions for future research are provided in Section 5.
2 DATA MI=
NING AND
MACHINE LEARNING
According=
to
In reinfo=
rcement
learning, an agent learns to perform a task within an environment. The
reinforcement learning agent has a repertoire or set of basic actions it can
execute, and at any given time, it is assumed to be "residing" in=
a
set of states. When the agent reaches the final state, the environment, a
teacher, or the agent itself provides a reward. Thus, most actions are not
rewarded, but rewards are given infrequently or "rarely" <=
/span>
2.1 Tradi=
tional
machine learning models
This arti=
cle
applied several ML algorithms for data classification: Decision Trees, Rand=
om
Forest, XGBoost, Naïve Bayes, Support Vector Machines (SVM), Logistic
Regression, and a combination of these methods (also known as “ensembles”)
using strategies such as voting, stacking, bagging, and boosting.
A classif=
ier
based on Decision Trees is structured as a tree-like algorithm similar to a
flowchart, where each internal node (non-leaf node) represents a test on an
attribute, each branch represents the outcome of the test, and each leaf no=
de
(or terminal node) contains a class label. The highest node in the tree is =
the
root node. The process of learning decision trees is performed using
class-labeled training tuples
According=
to
Inspired =
by
Bayes' theorem and the calculation of conditional probabilities, the method
estimates the label of a new record based on probability distributions
previously calculated using labeled data
According=
to
2.2 Ensemble me=
thods
For the b=
agging
strategy, the term "bagging" stands for "bootstrap
aggregating", where each training set is a sample with replacement, and
the aggregated classifier counts the votes and assigns the class with the
majority of votes to a new instance
A boosting
classifier is designed to produce a prediction rule by combining flexible
classifiers in sequence, generating a more powerful classifier based on the
adjusted weights of previous classifiers' performance
Commonly =
used,
the voting model is a process in which multiple learning techniques are
applied, or the same technique is used multiple times to create the base
classifiers, where each of these bases is trained with distinct data. This
process makes classification predictions, where the highest vote or score
assigned to a prediction is accepted
The ensem=
ble stacking
learning method consists of two phases: base classifier and meta-classifier=
2.3 Expla=
inable
model
This model
generates SHAP values that indicate the contribution of each attribute in a
specific sample, and the predictive model returns a projected output for ea=
ch
separate sample
3. MATERI=
ALS AND
METHODS
The data =
used in
this study were obtained from surveys conducted in 2020 and 2023 by the
Latinobarómetro Corporation, with both datasets undergoing preprocessing and
splitting into training and testing sets. Variables were empirically select=
ed
based on their relevance to the problem, ensuring their mutual presence in =
each
dataset. Consequently, 19 attributes were selected from each year's dataset,
with minimal differences between the years, such as accentuation of specific
names and classifications, which were subsequently unified during
preprocessing. The most significant challenge identified was that one varia=
ble
related to the respondent's country of origin lacked information for one of=
the
18 countries present in the 2020 data. Therefore, it was necessary to exclu=
de
this country to maintain consistency in the results of future analyses, thus
aligning the variables passed to the algorithms.
The selec=
ted
output variable was the respondent's assessment of their "general sati=
sfaction
with the economy" in their country, with all six different response
options present. To transform the problem into a binary classification, the
classes were grouped as 0 or 1 according to their correspondence, with 0
representing the "dissatisfied" group and 1 representing the
"satisfied" group. The transformation of the classes is presented=
in
Table 1.
Table 1
=
=
Tr=
eatment
of responses for the class variable
|
Classes from the original output |
New output |
|
Very satisfie=
d |
Satisfied (1)=
|
|
Somewhat satisfied |
|
|
Somewhat dissatisfied |
Insatisfied (=
0) |
|
Very dissatis=
fied |
|
|
Don’t know |
No cases
(excluded) |
|
No response |
To handle=
missing
data, no imputation methods were used, therefore, instances with incomplete
data were removed from both datasets. This strategy was chosen to maintain
greater reliability in the model training phase, while still retaining a
substantial amount of information even after excluding the missing data.
Additionally, attributes related to gender, country, race, and religion were
binarized, and the data were subsequently standardized and normalized. The =
data
processing and cleaning resulted in a final dataset consisting of 27,600
instances and 57 columns, with 14,032 instances for training and 13,568 for
testing.
Subsequen=
tly,
during the algorithm application, the preprocessed 2020 dataset was first u=
sed
as the training data, while the 2023 instances were used to test the models=
' effectiveness.
This approach allowed for the assessment of compatibility between the datas=
ets
from different years, ensuring that evaluating both years with the algorithm
would not affect the results, as the questions asked of the respondents
remained the same over a short period. After this step, the class balancing=
for
the training set reached 11,760 “dissatisfied” (83.81%) and 2,272 “satisfie=
d”
(16.19%).
For the
classification task, the methods Random Forest (RF), Logistic Regression (R=
EG),
Bernoulli Naïve Bayes (BNB), Support Vector Machine (SVM), XGBoost (XGB), a=
nd
Neural Networks (Neural) were used. To enhance these classifiers, the Grid
Search method was employed for hyperparameter tuning within a 5-fold
cross-validation and using accuracy as the reference metric. Table 2 summar=
izes
the best hyperparameters after tuning.
Table 2
=
The best parameters for each model after the Grid Search
|
Algorithm |
Chosen hyperparameters |
|
RF(n_=
estimators=3D100,
max_depth=3D20, min_samples_split=3D10, min_samples_leaf=3D1, max_feature=
s=3D'sqrt') |
|
|
REG(pe=
nalty=3D'l2',
C=3D0.001, solver=3D'liblinear') |
|
|
BNB(al=
pha=3D1.0,
binarize=3D1.0, fit_prior=3DTrue) |
|
|
SVM(C=
=3D1,
kernel=3D'rbf', gamma=3D'scale', max_iter=3D1000, probability=3DTrue) |
|
|
XGB(co=
lsample_bytree=3D1.0,
gamma=3D0.2, learning_rate=3D0.1, max_depth=3D3, n_estimators=3D100) |
|
|
Neural(hi=
dden_layer_sizes=3D(128,
), activations=3D'sigmoid', optimizer=3D'rmsprop') |
After tra=
ining
each traditional ML algorithm, ensemble methods were adopted among the
classifiers using the dataset that demonstrated the best accuracy. This
efficiency was evaluated based on the metrics of accuracy, precision, recal=
l,
and f1-score. Figure 1 presents the overall flowchart of the entire data
processing and model evaluation, implemented in Python programming language=
and
its libraries.
Figure 1
=
Research workflow
Finally, =
after
evaluating the performance of the classifiers and ensemble methods with the=
aid
of graphs, the SHAP library was applied to the algorithms that demonstrated=
the
highest quality. For this step, the Kernel Explainer function from the SHAP
library, with the algorithm being trained on the 2020 dataset and tested on=
the
2023 dataset, following the same method as the classifiers.
In the en=
d, this technique
provides a means to understand the decision-making method of the classifier,
elucidating the key factors of the highest-performing black-box model. This
enabled the testing and formulation of hypotheses regarding the categories =
that
most significantly influence public satisfaction or dissatisfaction with the
functioning of a country’s economy, cross-referencing these results with ar=
ticles
found in the literature.
4 RESULTS=
AND
DISCUSSION
The resul=
ts
obtained after applying the described methods were divided based on the
different types of criteria analyzed: (1) classifier algorithm and (2) ense=
mble
methods. Additionally, preference was given to presenting only the results
after applying the Grid Search method with all models already tuned to the
hyperparameters that maximized accuracy. Visualizations were provided for
better interpretations.
Initially=
, using
the prepared dataset as described in the previous section, all resulting
instances were employed in the classifier algorithms. To effectively
demonstrate the results achieved, all algorithms were tested with the same
input data, with the accuracy output revealing the best results, which are
analyzed in this section.
4.1 Perfo=
rmance
of the classifiers
Firstly, =
Figure 2
illustrates the performance of the classifiers without any ensemble method
applied, with the algorithm following the parameters and inputs previously
described. Thus, the heatmap information reflects the metrics used to measu=
re
the accuracy of each algorithm, with darker colors representing better resu=
lts,
or closer to the value of 1, and lighter colors indicating poorer classifie=
rs,
with metrics closer to a null value.
Figure 2
=
Classification results for individual algorithms
=
There was=
a
general difficulty among the algorithms in evaluating instances representing
people “satisfied” with the economy. All metrics for the "satisfied&qu=
ot;
block showed lower values compared to the other block. This result can be
explained by the imbalance in the training dataset, where instances of
dissatisfied people were more prevalent than their counterparts. The Random
Forest classifier had the best result in general, with a high f1-score for =
the “dissatisfied”
class, and 0.5 f1-score for “satisfied”.
Despite t=
his, it
is notable that among the instances, Logistic Regression presented the best
evaluation metrics for satisfied people, being considered, for this researc=
h,
the best classifier among the others. This result stems from its better bal=
ance
between instances, represented by its high precision rate for the satisfied
class while maintaining a higher f1-score and recall compared to the other
classifiers.
On the ot=
her
hand, the XGBoost method obtained the worst results, clearly affected by the
imbalance in the input instances, resulting in nearly null f1-score and rec=
all
figures for the minority class.
4.2 Perfo=
rmance
ensemble methods
From all =
the
classifiers previously used, while maintaining the same data input, the met=
hods
bagging, boosting, stacking, and voting (both in their hard and soft forms =
for
this last) were applied in search of an improvement in overall accuracy,
especially in the minority target variable. This type of model has consider=
able
potential for forming an algorithm with greater effectiveness in achieving
better results
Figure 3
=
Results for the ensemble algorithms
Evaluatin=
g the
average results of the ensemble algorithms in Figure 3, it is observed that=
the
models continued to perform better for the more favored class. For the
"Satisfied" class, the method that best managed to balance the
predictions was Stacking, with an f1-score of 56%. Regarding accuracies,
although the stacking method did not have the highest value for the
“dissatisfied” class, it demonstrated the best balance and is therefore
classified by this research as the method with the best overall results. It=
is
important to note that the results for “None” are related to the simple ave=
rage
of the individual classifiers as shown in Figure 2.
4.3 Featu=
re
analysis through explainable model
Even after
identifying a classifier algorithm with the best performance in the study, =
the
decision-making process indicating which factors most influenced the
determination of whether an individual was satisfied or dissatisfied with t=
he
economic situation remained unclear. To clarify the prediction method used =
by
the algorithm, the SHAP method was applied, as it is specifically designed =
to
better visualize the decision-making process of black-box models like some =
of those
used in this research. Given that, the stacking ensemble achieved the best
results, so SHAP was applied solely to this method to investigate the decis=
ion
factors.
The appli=
cation
of SHAP, following the parameters and procedures outlined in the previous
section, resulted in the graph presented in Figure 4. This figure displays =
the
names of the features that most significantly influenced the stacking metho=
d's
decision-making process, along with the values classified as influential for
determining satisfaction or dissatisfaction. Since not all instances exerte=
d a
strong influence on the decision-making process of this model, many of them
resulted in SHAP graphs without a defined SHAP value trend according to the
variable's value. These were therefore excluded from Figure 4, which includ=
es
only the six most important variables.
Figure 4
=
SHAP results obtained from the staking ensemble
First, is
noteworthy that much of the evaluation regarding economic performance is
related to the perception of other relevant aspects of human life. Accordin=
g to
the SHAP analysis, “satisfaction with democracy” is the most significant
factor, with individuals more content with their country's democracy tendin=
g to
also assess the economy more positively. This finding aligns with the
historical context of the continent, where maintaining democracy in a natio=
n is
highly correlated with its economic situation
Following
democracy, the assessment of the country's progress stands out, where
individuals who perceive their country as more prosperous tend to evaluate =
the
economy more favorably. Generally, the evaluation of this progress is
multifactorial, with studies confirming that indicators such as innovation =
The third=
most
important characteristic according to the SHAP analysis is a specific result
tied to just one nation. The graph in Figure 4 shows that Panamanian
respondents are more likely to positively evaluate their country's economy,
indicating that Panamanian citizens were more satisfied with their economy =
than
those in other Latin American countries. Although there are no direct studi=
es
on this relationship, the Inter-American Development Bank
The fourt=
h and
sixth factors, respectively, are satisfaction with life and perception of
personal financial improvement, highlighting a close relationship between
personal life quality factors and national economic performance. A related
study by The fifth
characteristic identified by SHAP as important for the stacking model is the
acceptance of authoritarian initiatives by the state if they solve societal
problems. Individuals more favorable to this type of governance demonstrate=
d a
greater likelihood of being economically satisfied. Although this finding
contrasts with the most significant factor (democracy), it was not consider=
ed
as confident as the other alternatives. The graph shows a large overlap of
positive and negative values for both classes of the interest variable.
Nevertheless, it suggests that the perception that an authoritarian regime =
can
solve societal problems remains strong in some Latin American countries due=
to
the historical instability of democratic regimes in the region. As seen f=
rom
their absence in both models, other characteristics of the respondents, suc=
h as
gender, age, religion, and race, did not significantly influence economic
satisfaction according to the SHAP investigation. These absences may indica=
te
that, despite the clear cultural differences among individuals from various
Latin American countries, economic satisfaction is defined by universal fac=
tors
that transcend community barriers. This conclusion is relevant as it sugges=
ts
that populations understand progress similarly, and sovereign states should
follow a similar path to better serve their citizens with a more advanced
economy. Finally, =
it was
found that the stacking ensemble model achieved the best results among the
models analyzed, especially when compared to the prediction made by individ=
ual
classifiers. This conclusion was based not only on the overall accuracy, wh=
ich
was lower than the other models but rather on the better balance between the
classes of the variable of interest, thereby providing results more aligned
with reality in assessing individuals satisfied and dissatisfied with their
country's economic situation. This factor is crucial because, with only high
accuracy, a model might favor instances of economic dissatisfaction simply
because they are the majority of recorded cases on the continent. Thus, the
stacking model effectively combined classifiers with low accuracy for the
minority variable, creating a new algorithm that better identified the actu=
al
satisfaction of individuals. On the other hand, the other ensemble models d=
id
not achieve satisfactory results, possibly due to data-related issues and t=
he
difficulty of predicting economic satisfaction when the proportion of outco=
mes
for the variable of interest indicates a significant rarity of such individ=
uals
in the surveyed population. 5 CONCLUS=
ION Based on =
the
results achieved by the interpretation model, this study identified several=
key
points relevant to the study of public perception of the economy. The strong
association found between the evaluation of democracy and economic satisfac=
tion
is well-supported by existing literature, confirming the study’s success in
reinforcing these findings, particularly in Latin America. Furthermo=
re, a
significant distinction was observed between traditional classifiers and
ensemble methods. Among the individual classifiers, Random Forest showed
superior performance, providing high accuracy for the “satisfied” class with
fewer samples, along with better balance among precision, and recall metric=
s.
This result is particularly relevant for future research involving classifi=
ers
for similar variables. Regarding
ensemble methods, the stacking model achieved the best results, surpassing =
the
Random Forest classifier in general. The observed outcome might be specific=
to
the dataset and variables analyzed, or due to characteristics of the chosen
classifiers. Further studies on the efficiency of these algorithms are
recommended to clarify these aspects. It is important to include data from =
past
years, seeking to have a higher training set, especially for the “satisfied”
class. In conclu=
sion,
this work met its objectives by comparing classifiers and addressing issues
related to public satisfaction with the economy in Latin America. Its
contributions are valuable for academics, professionals, policymakers, and
others interested in public economic perception studies, providing a
substantial resource for the field of computational intelligence. 6 ACKNOWL=
EDGMENTS The autho=
rs wish
to thank the Universidade Tecnológica Federal do Paraná for the Scientific
Initiation scholarship awarded to the second author of this work, through
Edital 02/2023 PROPPG – PIBIC, Cycle 2023-2024. REFERENCES Abdelka=
der,
H. E., Gad, A. G., Abohany, A. A., & Sorour, S. E. (2022). An Efficie=
nt
Data Mining Technique for Assessing Satisfaction Level With Online
Learning for Higher Education Students During the COVID-19. IEEE Access,=
10,
6286–6303. =
https://doi.org/10.1109/ACCESS.2022.3143035 Amin, M. N=
.,
Ahmad, W., Khan, K., Nazar, S., Arab, A. M. A., & Deifalla, A. F. (202=
3). Evaluating
the relevance of eggshell and glass powder for cement-=
based
materials using machine learning and SHapley Additive exPlana=
tions
(SHAP) analysis. Case
Studies in Construction Materials, 19, e02278. https://doi.org/10.1016/j.cscm.2=
023.e02278 Ariza, V. M. P., & Santos, =
B.
S. dos. (2023). Classificação da percepção de servidores públicos federais em
relação a atos de corrupção utilizando algorit=
mos
de aprendizado de máquina. Brazil=
ian
Journal of Production Enginee=
ring,
9(4), 166–178. https://doi.org/10.47456/bjpe.v9i4.42073<=
/a> Bah, I.=
A.
(2023). The relationship between education and economic growth: A cross=
-country
analysis. Research, Society and Development, 12(5), e19312540522. https://doi.org/10.33448/rsd-v12i5.40522<=
/a> Bortolotti, S. L. V., Moreira
Junior, F. de J., Bornia, A. C., Sousa Júnior, A. F. de, & Andrade=
, D.
F. de. (2012).=
Consumer
satisfaction and item
response theory: creatin=
g a
measurement scale. Gestão & Produção, 19(2),
287–302. https://doi.org=
/10.1590/S0104-530X2012000200005 Bozzetto, M., & Amador, R.
(2022). Experiências Institucionais e Confiança na Democracia:
Correlações entre as Instabilidades na América Latina. Congreso
Latinoamericano y Caribeño de Ciencias Sociales. “Democracia, Justicia e Igualdad,” 282–306.<=
o:p> Breiman, L. (2001). Random
forests. Machine
Learning, 45, 5–32. Brunton=
, S.
L., & Kutz, J. N. (2019). Data-Driven Science and Engineering. Cambrid=
ge
University Press. https://doi.org/10.1017/9781108380690=
Cahill,=
K.
E., McNamara, T. K., Pitt-Catsouphes, M., & Valcour, M. (2015). Linking
shifts in the national economy with changes in job satisfaction, employee
engagement and work–life balance. Journal of Behavioral and Experim=
ental
Economics, 56, 40–54. https://doi.org/10.1016/j.socec.2015.03.002 Chamorro-A=
talaya,
O., Arce-Santillan, D., Arévalo-Tuesta, J. A., Rodas Camacho,
L., Dávila-Laguna, R. F., Alejos-Ipanaque, R., & Moreno-Chinchay, L. R.
(2022). Supervised learning using support vector machine applied to sentime=
nt
analysis of teacher performance satisfaction. Indonesian Journal of
Electrical Engineering and Computer Science, 28(1), 516. https://doi.org/10.11591/ijeecs.v28.i1.pp516-524=
Da Silva, A. O., Raminelli, D. =
G.
de T. L., Dos Santos, B. S., & Lima, R. H. P. (2023). Classificação =
das
percepções de stakeholders sobre o futuro do Brasil utilizando aprendizad=
o de
máquina. AtoZ: Novas Práticas Em Informação e Conhecimento, 12, =
1. https://doi.org/10.5380/atoz.v12i0.84075 D’Ambro=
sio,
C., Frick, J. R., & Jäntti, M. (2009). Satisfaction with Life and
Economic Well-Being: Evidence
from Germany. Schmollers Jahrbuch, 129(2),
283–295. https://doi.org/10.3790/schm.129.2.283 Dandolo, D=
.,
Masiero, C., Carletti, M., Dalle Pezze, D., & Susto, G. A. (2023).=
AcME—Accelerated model-agnostic explanations: Fast whitening of the machine=
-learning
black box. Expert Systems with Applications, 214, 119115. https://doi.org/10.1016/j.eswa.2022.119115 Géron, =
A.
(2019). Hands-on: Machine Learning with Scikit-Learn, Keras & Tensorf=
low
(2nd ed.). O’Reilly Media. Graafla=
nd,
J., & Compen, B. (2012). Economic Freedom and Life Satisfaction: A Cross
Country Analysis. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.2057751 Han, J., P=
ei,
J., & Tong, H. (2022). Data
Mining: Concep=
ts
and Techniq=
ues(4th
ed., Vol. 1). Morgan Kaufmann. Inter-A=
merican
Development Bank. (2008). Idb | faster economic growth hurts life
satisfaction in latin america and the caribbean. Jafarza=
deh,
H., Mahdianpari, M., Gill, E., Mohammadimanesh, F., & Homayou=
ni,
S. (2021). Bagging and Boosting Ensemble Classifiers for Classif=
ication
of Multispectral, Hyperspectral and PolSAR Data: Compara=
tive
Evaluation. Remote Sensing, 13(21), 4405. https:/=
/doi.org/10.3390/rs13214405 Jiang, W.,
Chen, Z., Xiang, Y., Shao, D., Ma, L., & Zhang, J. (2019). SSEM: A Novel
Self-Adaptive Stacking Ensemble Model for Classification. IEEE Access,=
7,
120337–120349. https://doi.org/10.1109/ACCESS.2019.2933262 Joolfoo=
, K.
M. B. A., Jugurnauth, R. A., & Joolfoo, M. B. A. (2022). Applica=
tion
of Machine Learning in Predicting Customer Satisfaction of Telecom
Service Providers. 2022 4=
th
International Conference on Emerging Trends =
in
Electrical, Electronic and Communications Engineering (ELECOM), 1–10. https://doi.org=
/10.1109/ELECOM54934.2022.9965212 Kadkhodaei, H. R., Moghadam, A.=
M.
E., & Dehghan, M. (2020). HBoost: heterog=
eneous
ensemble classifier based on the Boosting method and entropy measure=
ment.
Expert Systems with Applications, 157, 113482. https://doi.org/10.1016/j.eswa.2020.113482 Kalita,=
J.
(2022). Machine Learning. Chapman and Hall/CRC. Khan, J.
(2015). The Role of Research and Development in Economic Growth. MPRA. Kowalsk=
i,
R. (2017). Patients’ written reviews as a resource for public healthc=
are
management in England. Proced=
ia
Computer Science, 113, 545–550. https://doi.org/https://doi.org/10.1016/j.procs.=
2017.08.275 Lan, H.,
Wang, S., & Zhang, W. (2024). Predicting types of human-related maritime
accidents with explanations using selective ensemble learning and SHAP
method. Heliyo=
n,
10(9), e30046. https://doi.org/10.1016/j.heliyon.2024.e30046 Langan,=
A.
M., & Harris, W. E. (2024). Metrics of student dissatisfaction and
disagreement:
longitudinal explorations of a national survey instrum=
ent.
Higher Education, 87(2), 249–269. https://doi.org/10.1007/s10734-023-01004-0 Li, Q.,
Jing, R., & Xihua Zhu. (2024). Determinants of travel satisfaction<=
o:p> for
commercial airlines: A data
mining approach. Engineering Applications of
Artificial Intelligence, 133, 108597. https://doi.org/10.1016/j.engappai.2024.108597 Liang, =
W.,
& Jia, C. (2023). Application of improved neighbor propagation algorit=
hm
in international communication and cooperation to promote interna=
tionalization
of higher education. Comput=
er
Applications in Enginee=
ring
Education, 31(3), 696–709. https://doi.org/10.1002/cae.22578 Martínez-N=
avalón,
J.-G., Gelashvili, V., & Gómez-Ortega, A. (2021). Evaluat=
ion
of User Satisfaction and Trust of Review Platforms: Analysis of the Imp=
act
of Privacy and E-WOM in the Case of TripAdvisor. Frontiers in Psychol=
ogy,
12. https://doi.org/10.3389/fpsyg.2021.750527 Naem, A.
A., Ghali, N. I., & Saleh, A. A. (2018). Antlion optimization and boosting
classifier for spam email detection. Future Computing and Informa=
tics
Journal, 3(2), 436–442. https:/=
/doi.org/10.1016/j.fcij.2018.11.006 Nipa, N=
.,
Riyad, M. H., Satu, S., Walliullah, Howlader, K. C., & Moni, M. A. (202=
4). Clinically
adaptable machine learning model to identify early appreci=
able
features of diabetes. Intell=
igent
Medicine, 4(1), 22–32. https:/=
/doi.org/10.1016/j.imed.2023.01.003 Noviant=
oro,
T., & Huang, J.-P. (2022). Investigating airline passenger satisfa=
ction: Data mining method. Research in
Transportation Business & Managem=
ent,
43, 100726. https://doi.org/10.1016/j.rtbm.2021.100726 Ogutu, =
R.
V. A., Rimiru, R. M., & Otieno, C. (2022). Target Sentiment Analysis
Ensemble for Product Review Classification. Journal of Information Technol=
ogy
Research, 15(1), 1–13. https://doi.org/10.4018/JITR.299382 Pecorar=
i,
N., & Cuesta, J. (2023). Citizen Participation and Political Trust in
Latin America and the Caribbean: A Machine Learning Approach. Policy
Research Wrking Papers, 10335. Polce, =
E.
M., Kunze, K. N., Fu, M. C., Garrigues, G. E., Forsythe, B., Nichols=
on,
G. P., Cole, B. J., & Verma, N. N. (2021). Development of supervi=
sed
machine learning algorithms for prediction of satisfaction at 2 years
following total shoulder arthroplasty. Journal of Shoulder and Elbow Surgery.
https://doi.org/10.1016/j.jse.2020.09.007 Rosa, D. M. de S., Dos Santos, =
B.
S., & Lima, R. H. P. (2023). Predicting satisfa=
ction
with democracy in Brazil considering data form an opinion survey.
Revista Gestão Da Produção Operações e Sistemas, 18. https://do=
i.org/10.15675/gepros.2965 Sabarmathi,
G., & Chinnaiyan, R. (2019). Reliable Machine Learning Approach to Pred=
ict
Patient Satisfaction for Optimal Decision Making and Quality Health
Care. 2019
International Conference on Communication and Electronics Systems
(ICCES), 1489–1493. https://doi.org/10.1109/ICCES45898.2019.9002593 Sagi, O=
.,
& Rokach, L. (2018). Ensemble learning: A survey. WIREs Data Mining =
and Knowledge
Discovery, 8(4). https://doi.org/10.1002/widm.1249 Tauil, Y. B., Santos, B. S. dos,
& Lima, R. H. P. (2024). Machine learning techniq=
ues
in classifying satisfaction with the economy of Latin American Citizens.<=
/span> Observatorio de La Economía
Latinoamericana, 22(5), e4912. https://do=
i.org/10.55905/oelv22n5-199 Van Ryzin,=
G.
G. (2004). The
Measurement of Overall Citizen Satisfaction. Public
Performance & Management Review, 27(3), 9–28. http://www.jstor.org/stable/3381143 VeČ=
;erník,
J., & Mysíková, M. (2015). GDP and life satisfaction in European countri=
es –
focus on transition.
Post-Communist Economies, 27(2), 170–187. https:/=
/doi.org/10.1080/14631377.2015.1026687 Veenhov=
en,
R. (1996). Developments in satisfaction-research. Social Indicat=
ors
Research, 37(1), 1–46. https://doi.org/10.1007/BF00300268 Vishwar=
upe,
V., Joshi, P. M., Mathias, N., Maheshwari, S., Mhaisalkar, S., &
Pawar, V. (2022). Explainable AI and Interpretable Machine Learning: A Case St=
udy
in Perspective. Procedia Computer Science, 204, 869–876. https:/=
/doi.org/10.1016/j.procs.2022.08.105 Yadav, =
V.,
Dubey, A. K., Singh, H. P., Dubey, G., & Suryani, E. (2022). Process
Mining Techniques for Pattern Recognition. CRC Press. https:/=
/doi.org/10.1201/9781003169550 Zaghlou=
l,
M., Barakat, S., & Rezk, A. (2024). Predicting E-commerce customer
satisfaction: Traditi=
onal
machine learning vs. deep learning approac=
hes.
Journal of Retailing and Consumer Services, 79, 103865. https:/=
/doi.org/10.1016/j.jretconser.2024.103865 Zhang, =
S.,
Chen, J. Y., Pang, H. N., Lo, N. N., Yeo, S. J., & Liow, M. H. L. (202=
1). Development
and internal validation of machine learning algorit=
hms
to predict patient satisfaction after total hip arthroplasty. Arthrop=
lasty,
3(1), 33. https://doi.org/10.1186/s42836-021-00087-3 Zheng, =
X.,
Xie, Y., Yang, X., Amin, M. N., Nazar, S., Khan, S. A., Althoey, F., &am=
p;
Deifalla, A. F. (2023). A data-driven approach to predict the compres=
sive
strength of alkali-activated materials and correlation of influen=
cing
parameters using SHapley Additive exPlanations (SHAP) analysis. Journal=
of
Materials Research and Technology, 25, 4074–4093. https:/=
/doi.org/10.1016/j.jmrt.2023.06.207 Zhylins=
ka,
O., Chernyak, O., & Bazhenova, O. (2019). The role of innovat=
ions
in driving economic growth: case of advanced economies. Globali=
zation
and Business, 4(7), 11–15. https:/=
/doi.org/10.35945/gb.2019.07.001 Žižka, =
J.,
Dařena, F., & Svoboda, A. (2019). Text mining with machine learnin=
g: Principles and Techniques (1st
ed.). CRCPress. Zou, Z.,
Wang, L., Chen, J., Long, T., Wu, Q., & Zhou, M. (2022). Research on pean=
ut
variety classification based on hyperspectral image. Food Science and
Technology, 42. https://doi.org/10.1590/fst.18522
Análise da satisf=
ação
com a economia a partir de modelos de aprendizado de máquina e inteligência
artificial
explicativa
Luiz Fernando Menegazzo
Ferreyra; Yasser Bulaty Tauil; Helton Messias Adigneri; Bruno Samways dos
Santos; Rafael Lima
|
IS=
SN
2237-4558 • Navus
• Florianópolis •
SC • v. 15 • p. 01-17 • jan./dez. 2024 |
|
|
ISSN 2237-4558
•
Navus •
Florianópolis • SC |
|