DIAGNOSIS OF OBESITY LEVEL BASED ON BAGGING ENSEMBLE CLASSIFIER AND FEATURE SELECTION METHODS

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.13, No.2, March 2022
DOI: 10.5121/ijaia.2022.13203 37
DIAGNOSIS OF OBESITY LEVEL BASED ON
BAGGING ENSEMBLE CLASSIFIER AND
FEATURE SELECTION METHODS
Asaad Alzayed, Waheeda Almayyan and Ahmed Al-Hunaiyyan
Computer Information Department, Collage of Business Studies, PAAET, Kuwait
ABSTRACT
In the current era, the amount of data generated from various device sources and business transactions is
rising exponentially, and the current machine learning techniques are not feasible for handling the massive
volume of data. Two commonly adopted schemes exist to solve such issues scaling up the data mining
algorithms and data reduction. Scaling the data mining algorithms is not the best way, but data reduction
is feasible. There are two approaches to reducing datasets selecting an optimal subset of features from the
initial dataset or eliminating those that contribute less information. Overweight and obesity are increasing
worldwide, and forecasting future overweight or obesity could help intervention. Our primary objective is
to find the optimal subset of features to diagnose obesity. This article proposes adapting a bagging
algorithm based on filter-based feature selection to improve the prediction accuracy of obesity with a
minimal number of feature subsets. We utilized several machine learning algorithms for classifying the
obesity classes and several filter feature selection methods to maximize the classifier accuracy. Based on
the results of experiments, Pairwise Consistency and Pairwise Correlation techniques are shown to be
promising tools for feature selection in respect of the quality of obtained feature subset and computation
efficiency. Analyzing the results obtained from the original and modified datasets has improved the
classification accuracy and established a relationship between obesity/overweight and common risk factors
such as weight, age, and physical activity patterns.
KEYWORDS
Data mining, Obesity, Feature reduction, Filter feature selection, Bagging algorithm.
1. INTRODUCTION
The World Health Organization defines obesity and overweight as excessive fat accumulation in
particular body parts that can be dangerous to wellbeing. Obesity among adults, teens, and
children has emerged worldwide health problem. The number of people who suffer from obesity
has doubled since 1980. The studies estimated that more than 1900 million adults nowadays
suffer from weight alteration, and by 2030, lifestyle diseases will contribute to 30% of global
deaths [1]. There is an alarming increase in obesity, and overweight is the primary lifestyle
disease that leads to other health disorders, such as cardiovascular diseases, chronic obstructive
pulmonary disease, cancer, type II diabetes, hypertension, anxiety, and depression [2].
Obesity is a disease with numerous factors, such as uncontrolled weight gain, excessive fat
intake, and energy consumption. Some of the reasons for being overweight are the increased
consumption of high energy-dense food high in fat and ow frequency of physical activity due to
an inactive type of work, the new transport manners, and growing urbanization. Biological hazard
factors such as genetic background can also cause obesity, so there can be several kinds of
obesity as Monogenic, leptin, polygenic and syndromic. Besides, other risk factors are social,

38
psychological, and unhealthy eating habits [3]. On the other side, some literature suggests other
determining several potential risk factors that contribute to obesity, such as being an only child
family conflicts as divorce, depression, and anxiety [4,5]. Identifying risk factors can prevent
obesity and overweight with the appropriate behavioral intervention strategies [6].
Data mining is the analytical process of knowledge discovery in large and complex datasets,
obtaining new information to support decision-making [7,8,9]. This study had the objective of
implementing several data mining techniques to predict overweight/obesity status. Several
scholars have looked into this disorder and generated means to figure out the obesity level of an
individual; yet, such tools are limited to calculating the BMI, neglecting relevant factors such as
the family history of obesity and leisure time dedicated to physical activities [9]. We considered
building a competent tool to detect obesity levels more efficiently based on this.
Many lifestyle factors or unhealthy habits contributing to obesity have been reported, ranging
from dietary and physical activity patterns [2, 4, 5, 6]—Accordingly, an apparent demand to
develop an automated solution to diagnose obesity. Consequently, a large scale of datasets is a
significant challenge. The feature selection process minimizes the computational overhead and
improves the overall performance by eliminating nonessential and unimportant features from
datasets before model implementation [10], a crucial requirement in medical data classification
problems. A considerable number of machine learning algorithms have been proposed for
medical datasets and related classification problems, including neural networks methods
[11,12,19], decision trees algorithms [11,14,19], convolutional neural network algorithm
approaches [22,23], SVM model approaches [13,16,17], and k-nearest neighbor classifiers
[13,18].
Obesity already exists; this article seeks to explore the critical factors behind this disease from a
data-mining point of view. The primary objective is to improve the prediction accuracy of obesity
with a minimal number of feature subsets using the dataset gathered by Palechor and de la Hoz
Manotas to estimate obesity levels among people from Mexico, Peru, and Colombia [24]. The
main contributions of this paper include four aspects.
I. In the current work, we investigate the performance of different feature selection
methods to build optimal feature subsets to consider the obesity and overweight risks
such as The Inconsistency Metric, mRMR, ReliefF, Pairwise Consistency, Pairwise
Correlation, and Tabu search.
II. We added Basal Metabolic Rate, Resting Metabolic Rate, and Body Fat Percentage
parameters to the dataset list of features.
III. We built several machine learning classifiers, C4.5, FURIA, RandF, and Bagging, to
classify the feature subsets.
IV. We carried out experiments considering the 10-fold cross-validation and several
evaluation metrics to test the effectiveness of the suggested classification algorithm in
predicting the obesity level.
The remainder of the article is as follows; Section 2 consists of the existing literature on using the
machine learning model in overweight and obesity prediction. Then Section 3 represents the
approaches used in this article. Section 4 clarifies the methodology that includes the proposed
feature selection technique and the classifiers. Section 5 discusses the detail regarding
implementing the proposed classification algorithm and details about experiments and results.
Finally, conclusions are given in Section 6.

39
2. PREVIOUS WORK
Recently, data mining has been vastly applied in many fields, including healthcare and medical
science. Several data mining techniques have been used to explore the obesity problem among
adolescents. In [11], researchers evaluated different multivariate regression methods and
multilayer perceptron feedforward neural network models to define people at risk of obesity
using the UK Millennium Cohort Study set up to follow the lives of children born at the turn of
the new century. With an accuracy of above 90% to predict teenager BMI from prior BMI
readings. In a recent research [12] that carried on the experiments over the dataset, they used the
Synthetic Minority Oversampling Technique (SMOTE) technique to handle the imbalance issue
and ensemble classifier.
Bassam et al. [13] built analytical models on data obtained from the Kuwait Health Network to
provide early insight into the future risk of type II diabetes. The study considered several
parameters, including age, gender, BMI, pre-existing hypertension, and family history of
hypertension. Using a machine-learning algorithm that compares the performance of logistic
regression, k-nearest neighbor (kNN), support vector machine (SVM) based on a five-fold cross-
validation technique, the kNN classifier has outperformed the other classifiers in terms of AUC
values. Meghana et al. [14] developed a machine learning tool using auto-sklearn model to
classify two cardiovascular disease datasets and compare the classification accuracies by the
opinion of graduate students. The results reported outperformed traditional machine learning
classifiers and the graduate student.
Jindal et al. [15] conducted a research R ensemble prediction model and Python interface model
for obesity prediction based on age, height, weight, and BMI. The ensemble approach scored an
accuracy of 89.68% and outperformed generalized linear model, random forest, and partial least
squares. Seyla et al. [16] investigated the impact of dietary and physical activity behaviors on
obesity using discriminant analysis, support vector machines (SVM), and neural nets algorithms.
As a result, SVM achieved a higher prediction accuracy and validated that dietary and exercise
behavior patterns can better explain the prevalence of obesity instead of individual components.
Dunstan et al. [17] utilized SVM, Random Forest, and Extreme Gradient Boosting to predict
country-level obesity prevalence, based on countrywide food sales of a small subset of food and
beverage group categories. Random Forest predicted obesity with an absolute error of 10%. The
study indicated that commercial baked goods and flours, followed by cheese and sweet
carbonated drinks, were the most appropriate food categories in predicting obesity. Zheng et al.
[18] analyzed the 2015 Youth Risk Behavior Surveillance System dataset for the state of
Tennessee in order to predict obesity in high school students by focusing on nine health-related
behaviors. They have utilized binary logistic regression, improved decision tree (IDT), weighted
kNN, and neural network. Results showed that the weighted kNN model achieved an 88.82%
accuracy and 93.44% specificity in the classification problem.
DeGregory et al. [19] proposed adapting smart wearable sensor devices, electronic healthcare
records, and smartphone apps in obesity-related data gathering to prevent obesity/overweight
problems. They studied the behavior of several machine learning methods and topological data
analysis on the National Health and Nutrition Examination Survey. Golino et al. [20] used a
random tree technique to investigate the prediction of elevated blood pressure using BMI, waist
and hip circumference, and waist-hip ratio to study data collected from 400 students. The
proposed model outperformed the traditional logistic regression model in predictive power. The
proposed model reported a sensitivity of 80.86%, specificity of 81.22% in the training set, and
45.65%, 65.15% in the test sample for women. Moreover, a sensitivity of 72% and specificity of
86.25% in the training set and, respectively, 58.38% 69.70% in the test set for men.

40
Pleuss et al. [21] led research to deploy a machine learning-based model to process a 3D image to
obtain hundreds of anthropometric measurements after analyzing the images obtained from a 3D
scanner. They used anthropometric information to estimate BMI, waist circumference, and hip
circumference to body fat. Maharana et al. [22] built a regression model based on a convolutional
neural network to approximately process 150,000 3-D satellite images from Google Static Maps
API in six cities in the US to extract features of the built environment to study connections
between the built environment and obesity. They concluded that consistently presents a strong
association between obesity prevalence and the built environment indicators, despite varying city
and neighborhood values. Pouladzadeh et al. [23] proposed a combination of graph cut
segmentation and deep learning neural networks to classify and recognize food items that would
run on smartphones to calculate the amount of calorie intake automatically. Results showed that
combining the two methods provides 100 % food recognition accuracy. From the related work,
we identified a list of machine learning models and risk factors related to obesity/overweight, as
described in Table 1.
Table 1. The risk factors related to obesity, overweight according to data mining approaches
Researcher Risk Factors
DeGregory et al. [19] Inactivity, inappropriate diet
Singh et al. [11,12] BMI
Bassam et al. [13] Age, gender, BMI, pre-existing hypertension, family history of
hypertension, and diabetes type II
Meghana et al. [14] Cardiovascular diseases
Seyla et al. [16] Activity, nutrition
Jindal et al. [15] Age, height, weight, BMI
Zheng et al. [18] Inactivity, improper diet
Dunstan et al. [17] Unhealthy diet
Golino et al. [20] Blood Pressure, BMI, Waist Circumference, Hip Circumference,
Waist–Hip Ratio
Pleuss et al. [21] BMI, Circumference, Hip Circumference
Maharana et al. [22] Environment, context
Pouladzadeh et al. [23] Nutrition
3. METHODOLOGY AND MATERIALS
The primary objective of this research work is to study the performance of various machine
learning classifiers with various feature sets. One of the motivations in this research is to improve
classification performance by applying several feature selection approaches to build a new
feature subset and several machine learning classifiers for prediction. The process can be
completed in two steps: feature selection and classification. In the framework of the proposed
approach, we utilized the Bagging algorithm to predict the obesity level. The Bagging algorithm
has been used as an effective ensemble algorithm to ensure good performance of the base
classifiers in this work since the Bagging algorithm can thoroughly handle the core characteristic
of base classifiers algorithms, instability [19]. It also utilizes the Bagging algorithm as a reliable
method due to its generalized capability to avoid overfitting problems by reducing the variance.
Generally, the generalization capability of the classification model is reduced by the existence of
redundant features. Therefore, we apply the Filter-based feature selection method to construct a
compact feature selection of high-impact features and reduce redundant features.

41
3.1. Dataset
Overweight and obesity are related to various factors, such as food and dietary practices, energy
consumption, genetics, lifestyle, socioeconomic aspects, anxiety, and depression [3]. The
deployed dataset was constructed by [24] after searching for literary sources for the main factors
or habits contributing to obesity. The sample population included 1043 females and 1068 males
between 14 and 61 years old, including Colombia, Mexico, and Perú. The dataset includes 2111
records and 18 variables based on the surveys to identify their obesity level. Information gathered
included age, gender, weight, the frequency of physical activity, the frequency of fast food intake,
and other aspects that could help describe the nutritional behavior of obese people (see Table 2).
To specify the obesity levels, we used the table provided by WHO (Table 3) to categorize
correctly the data analyzed based on the BMI.
Most of the literature concerning obesity depends on BMI to assess overweight and obesity,
although it is not generic enough to be applied in every context, such as a pregnant lady or an
older man. Therefore, we decided to include metabolic and anthropometric measures. We decided
to add Basal Metabolic Rate, Resting Metabolic Rate, and Body Fat Percentage parameters to the
dataset.
Basal metabolic rate (BMR) defines the rate of energy consumed by the human body. According
to the Harris-Benedict equation [25],
1
Resting metabolic rate (RMR) depends on age, weight, height, gender. According to [26],
2
Body fat percentage (BFP) and according to Deurenberg [27] If the person is a child, then
3
Table 2. Collected dataset description
Attribute Abbreviation Type Possible values
1. Gender - Nominal Male, Female
2. Age - Numeric Integer Numeric Values
3. Height - Numeric Integer Numeric Values
(Mt)
4. Weight - Numeric Integer Numeric Values
(Kg)
5. Family with
overweight/Obesity
FHWO Nominal Yes, No
6. Fast Food Intake FAVC Nominal Yes, No
7. Vegetables Consumption
Frequency
FCVC Numeric 1: Always
2: Sometimes
3: Rarely
8. Number of main meals
daily
NCP Numeric 1 to 2: UD 3: TR More
than 3: MT

42
9. Food intake between
meals
CAEC Nominal S: Always
CS: Usually
A: Sometimes
CN: Rarely
10. Smoking SMOKE Nominal Yes, No
11. Liquid intake daily CH2O Numeric 1: Less than one-liter
2: Between 1- and 2-liters
3: More than 2 liters
12. Calories Consumption
Calculation
SCC Nominal Yes, No
13. Physical Activity FAF Numeric 1: 1 to 2 days
2: 3 to 4 days
3: 5 to 6 days
0: No physical activity
14. Schedule dedicated to
technology
TUE Numeric 0: 0 to 2 hours
1: 3 to 5 hours
2: More than 5 hours
15. Alcohol consumption CALC Nominal NO: No consume of
alcohol
CF: Rarely
S: Weekly D: Daily
16. Type of Transportation
used
MTRANS Nominal TP: Public transportation
MTA: Motorbike
BTA: Bike
CA: Walking
AU: Automobile
17. Class IMC Nominal Based on the WHO
Classification:
-Underweight
-Normal
-Overweight
-Obesity I
-Obesity II
-Obesity III
Table 3. Definition of underweight, overweight and obesity according to the WHO reference system for
México [24]
WHO Category BMI (kg/m)
Underweight Less than 18.5
Normal 18.5-24.9
Overweight 25.0-29.9
Obesity Class I 30.0-34.9.
Obesity Class II 35.0-39.9
Obesity Class III More than 40.0
3.2. Feature selection
The feature selection phase aims to specify essential attributes before constructing a classification
model by removing non-essential and redundant attributes. There are two main feature selection
paradigms: Filter and Wrapper methods, where each method has a unique feature selection
mechanism [28]. The filter methods evaluate the feature's relevancy by utilizing a ranking
procedure that withdraws the most minor ranked features accordingly. This method enhances the
correlation between features and classes through the evaluation criteria and weakens the

43
correlation between features. Due to that, the Filter method is robust, scalable, computationally
efficient, and most importantly, independent of the classifier.
On the other hand, the wrapper model works nearly like the filter methods, but they evaluate the
subsets using a predefined classification algorithm rather than an independent measure.
Correspondingly, the predictive accuracy is often high, while the generalization ability is
inadequate. Therefore, the wrapper methods deliver more satisfactory results, but they tend to be
more computationally expensive with large-scale datasets [29].
Overall, we can summarise the difference between the Filter and Wrapper methods. The Filter
method has a better-generalized ability, though the estimation performance of the Wrapper
method is much better. Moreover, the computational cost of the Wrapper method is better, and it
has a more significant potential to handle the overfitting issue. Therefore, several articles have
applied feature selection methods before conducting the classification process [12, 18, 20].
1. Tabu Search Technique
Tabu search is an iterative memory-based algorithm proposed by Glover in 1986 to solve
combinatorial optimization problems [30]. It contains a local search mechanism conjoined with a
tabu mechanism. Tabu search starts with an initial solution X’ among neighborhood
solutions, where  feasible solutions are set. Since then, several researchers have applied the
Tabu successfully search in several multiclass classification problems [31].
The algorithm explores and assesses all the possible solutions N(X)  to obtain a new one, X’
 N(X), with a better functional value if the new solution X’ is not listed in the tabu list or it
satisfies the aspiration criterion [24]. If the candidate solution 𝑋’ is better than 𝑋best, the value of
𝑋best is overridden; else, the Tabu search will go uphill to avoid local minima. Moreover, to avoid
local minima, Tabu search limits visiting previously visited solutions for a certain number of
iterations. Next, the neighborhood search resumes the new solution 𝑋’ until it meets the stopping
criterion.
2. mRMR feature selection method
The mRMR is a heuristic technique proposed by Peng et al. to measure the relevancy and
redundancy of features and determine the most informative features [32]. The mRMR feature
selection technique aims to simultaneously identify features with maximum relevancy for target
classes and minimum redundancy with other features in the dataset. The mRMR method helps
determine crucial features and, at the same time, minimize the classification error. The basic
concept of the mRMR method is to use two mutual information operations: one to measure the
relevancy between classes and each feature and the second to measure redundancy between every
feature. The main goal of applying the mRMR feature is to find a subset of features from 𝑆 with
𝑚 features, {𝑥𝑖}, that jointly has the most considerable dependency on the target class C or have
the minimal redundancy on the selected features subset 𝑆.
3. ReliefF
ReliefF is a well-known multivariate filter that extends a prior version of Relief [33] based on
nearest neighbors. It randomly selects samples and searches for nearest neighbors from the same
class (ignoring the rest). The ReliefF works by comparing the values of the selected sample with
the hits and misses, and then it updates the relevance score for each feature. A helpful feature
should weigh similar examples from the identical class and different instances from the other

44
classes similar to examples from the same class and different from examples from the other
classes.
4. The Inconsistency Metric
The idea behind this attribute subset evaluation method is to find a subset of features that can
separate the dataset with a highly predominant class [34]. The consistency measure approach
finds out the best discrimination set from the original data rather than applying an algorithm to
split the classes. The inconsistency measure considers the features inconsistent when two or more
samples have the same values but distinct labels.
5. Pairwise Correlation
The Pairwise Correlation technique was developed by Jiménez F. et al. in 2021 and was inspired
by the correlation-based feature selection method [35]. The proposed Pairwise Correlation
method evaluates an attribute i by using the following function
4
where ΦD ({i, j}) is the merit of the subset formed by attributes i and j, for all j = 1, . . ., n,
5
The merit ΦA
D (i) of an attribute i is the mean of the merits ΦD of the attribute subsets formed
by i and each of the other attributes. The method prefers attributes with low correlation to other
attributes, and the class is highly correlated.
6. Pairwise Consistency
The Pairwise Consistency is a relatively new feature selection method introduced by Jiménez F.
et al. that developed the consistency metric for attribute subsets introduced by Liu and Setiono
[35, 36]. The idea behind the consistency measure is to locate attributes that can split the dataset
into parts with a favorably predominant class. The proposed Pairwise Consistency method
evaluates an attribute i ∈ {1, . . ., n} by using the following function Ψ AD
6
where ΨD({i,j}) = 1 − ID({i,j}) is the consistency rate of the subset formed by the attributes i and
j, for all j = 1, . . . , n, with j‡i. The merit ΨAD (i) of an attribute i is the mean of the consistency
rates of the attribute i and each of the other attributes.
4. CLASSIFICATION
4.1. C4.5 decision tree Classifier
C4.5 is an improved version of the ID3 decision tree algorithm. Its capabilities include
approximating discrete-valued functions, robust noisy data, and handling missing attribute values.
The C4.5 algorithm chooses the best splitting attributes using the information gain (IG) ratio as
the default attribute choice criterion [37]. This algorithm begins visiting each node to select an

45
optimal split based on the IG ratio's maximization and designates it as the root node of a tree.
Then, it forms a leaf for all of the potential values. The algorithm's primary mission is to choose
the appropriate feature to partition the dataset into several categories.
4.2. RandF (RandF) Machine Learning
RandF algorithm is a tree-based-ensemble machine learning method based on a combination tree
of predictors. Each tree uses a random vector sampled independently from the classification input
vector. Each tree depends on the values of a random vector sampled independently and with the
same distribution for all trees [38]. The algorithm assesses the last decision by compiling the
votes from all the trees using a rule-based approach or an iterative error minimization technique
by reducing the weights for the correctly classified samples. RandF accelerates the building speed
by building a variety of individual base classifiers.
4.3. Fuzzy Unordered Rules Induction (FURIA) classifier
Hühn and Hüllermeier first introduced the FURIA algorithm to improve and extend the state-of-
the-art rule learner algorithm RIPPER. FURIA is more accurate than the original RIPPER
algorithm and C4.5 classifier [39]. FURIA depends on fuzzy rules and unordered rule sets instead
of conventional rules for classification. It also uses the rule stretching technique to deal with
uncovered cases to generate an unordered one vs. rest scheme for each class. For that, the order
of the rule is not considered because is no default rule as each class is separated from others.
FURIA applies a rule stretching strategy if any rule may not cover a new query. Fuzzy intervals,
namely fuzzy sets with trapezoidal membership functions, are used to obtain fuzzy rules.
4.4. Bagging classifier
Breiman proposed bootstrap aggregating bagging as a meta-algorithm based on decision trees'
ensemble to improve classification and regression models [40]. Although this technique can be
applied in any algorithm, it mainly applies to decision tree models. Moreover, bagging helps
decrease the variance and reduce the over-fitting of estimation. Bagging diversity is achieved by
generating copies of the original training set, where different training datasets are randomly
drawn with replacement. Accordingly, a decision tree is built based on the standard approach
with each training data replica. Consequently, every tree can be described by a distinctive set of
variables, nodes, and leaves. Ultimately, their projections are joint to obtain the outcome [41].
4.5. Evaluation Metrics
The performance of the proposed model is measured using precision, the receiver operator
characteristic (ROC), and Cohen's kappa coefficient. We can define precision as
precision= 7
FP is the false positive rate, and TN is the True negative rate.
The ROC curve is a standard metric for analyzing classifier performance over a range of trade-
offs between true positive rate (TP) and false-positive rate (FP) [42]. ROC usually ranges from
0.5 for a perfectly random classifier and 1.0 for a perfect classifier. The area under the ROC
curve is a graphical plot for evaluating two-class decision problems.

46
Kappa error or Cohen's kappa coefficient is a useful measure to compare different classifiers'
performance and selected features' quality ranges from 1 to -1 [42].
If the kappa value -according to the successive formula- approaches +1, there is a better
opportunity for an agreement, and when it approaches -1, it indicates a poor chance for
agreement.
Kappa error = = 8
P(A) is the total agreement probability, and P(E) is the theoretical probability of chance
agreement.
5. RESULTS AND DISCUSSION
Since we have defined the theoretical mechanism of the proposed algorithm, it is necessary to
examine its practical efficiency as well. Accordingly, we plan to conduct several simulation
experiments to verify the reliability of the proposed algorithm. In the simulation experiments, we
have chosen to apply the 10-fold cross-validation method for the validation, and separately
recorded classification accuracy, AUC, Kappa, and MCC of FURIA, C4.5, RandF, and Bagging
classifiers. Initially, we should report the classifications without feature selection to demonstrate
the performance of the suggested feature selection techniques. Table 2 reports the performance
before performing feature selection. The RandF classifier (95.78%) slightly outperformed the
FURIA (95.07%) and C4.5 (93.74%) classifiers. At the same time, the Bagging based on the
RandF classifier outperformed the other classifiers by scoring accuracy of 95.878%. In most
classifiers ' performances, the results showed good performance in AUC, Kappa, and RMSE
metrics. The time to build the model is higher in FURIA and bagging-based FURIA than C4.5
and RandF classifiers.
Table 4. Classification performance without feature selection methods
Classifier ACC(%) AUC Kappa RMSE Time(s)
FURIA 95.07 0.987 0.9425 0.1103 1.85
C4.5 93.74 0.978 0.927 0.1286 0.02
RandF 95.78 0.998 0.9508 0.1143 0.44
Bagging(FURIA) 96.7314 0.998 0.9618 0.0938 10.79
Bagging(C4.5) 95.405 0.997 0.9463 0.1039 0.14
Bagging(RandF) 95.878 0.998 0.9519 0.1202 2.89
Table 4 lists the obtained features using mRMR, ReliefF, Pairwise Consistency, Pairwise
Correlation, The Inconsistency Metric, and Tabu search. This step has reduced the feature size
from 16 to 11 and 7. The feature selection methods helped reduce the number of attributes of the
original datasets, notably the Inconsistency Metric and TabuSearch techniques. Figure ?? shows
the resulting Venn diagram, which demonstrates the relationship among Feature selection
Techniques. Table 5 illustrates all the possible intersections between Feature selection
Techniques using the original dataset. All the feature selection techniques share Weight, CH2O,
and TUE attributes. Most of the methods share Age, FAF, and Height attributes.

47
Table 5. Optimal feature subsets
Feature Selection Method No. of Features Features
1. mRMR 10 Gender, Age, Weight, FHWO, FAVC,
FCVC, NCP, CAEC, CH2O, TUE
2. ReliefF 11 Gender, Weight, FHWO, FCVC, CALC,
CAEC, TUE, MTRANS, Height, NCP,
CH2O
3. Pairwise Consistency 9 Weight, Age, FCVC, FAF, Height, CH2O,
TUE, NCP, Gender
4. Pairwise Correlation 8 Weight, FCVC, Age, TUE, NCP, Gender,
CH2O, FAF
5. The Inconsistency
Metric
7 Age, Height, Weight, FAVC, CH2O, FAF,
TUE
6. TabuSearch 7 Age, Height, Weight, CAEC, CH2O, FAF,
TUE
Figure 1. Venn diagram demonstrating the cross-links between the feature selection techniques used over
the original dataset
Table 6. Intersecting Features among Feature selection Techniques using the original dataset
Feature Selection Techniques Intersecting Features
[mRMR ] and [ReliefF] and [PairwiseCons.] and [PairwiseCorr.] and
[Inconsistency.M] and [TabuSearch]
Weight, CH2O, TUE
[mRMR ] and [PairwiseCons.] and [PairwiseCorr.] and
Age
[PairwiseCons.] and [PairwiseCorr.] and [Inconsistency.M] and
[TabuSearch]
FAF
[ReliefF] and [PairwiseCons.] and [Inconsistency.M] and [TabuSearch] Height
[mRMR ] and [ReliefF] and [TabuSearch] CAEC
[mRMR ] and [Inconsistency.M] FAVC
[mRMR ] and [ReliefF] and [PairwiseCons.] and [PairwiseCorr.] Gender, FCVC, NCP
Table 6 reports the classification performance after performing feature selection. The RandF
classifier outperformed the other classifiers achieving an accuracy of 96.77% using the Pairwise
Consistency that yielded nine features. At the same time, the bagging based on the FURIA
algorithm outperformed the other classifiers by scoring an accuracy of 96.73% while using the
Pairwise Consistency technique. As the feature selection methods reduced the features of

48
datasets, it also increased the overall performance accuracy, as with the ReliefF, Pairwise
Consistency, The Inconsistency Metric, and TabuSearch techniques. Although the number of
features chosen by both the Inconsistency Metric and TabuSearch techniques is less than the
other techniques such as mRMR and Pairwise Correlation, the classification results were better.
The Bagging algorithm enhanced the prediction accuracy. Overall, the performance of the
Bagging algorithm is more promising than that of the base algorithms independently. It is worth
noting that the construction time for the model is reduced, chiefly for ReliefF, Pairwise
Consistency, Inconsistency Metric, and TabuSearch techniques.
All in all, it indicates that the feature selection step successfully enhances the classification
accuracy. Meanwhile, the Inconsistency Metric and Tabu search helped improve the
classification performance with a limited number of features. All the suggested models had AUC
values above 0.95. Kappa values were between 0.85 and 0.96. the values of RMSE slightly above
0.1. These results demonstrate that the models were successful in predicting obesity. However,
the time taken to build the model of FURIA and FURIA bagging-based classifiers is longer than
the other classifiers.
Table 7. Classification performance after feature selection
1. mRMR:10
Classifier ACC AUC Kappa RMSE Time
FURIA 87.873 0.962 0.8583 0.1656 1.69
C4.5 87.8257 0.962 0.8583 0.1656 0.08
RandF 92.3259 0.993 0.9104 0.133 0.42
Bagging(FURIA) 91.568 0.991 0.9015 0.1372 13.31
Bagging(C4.5) 88.9626 0.986 0.8711 0.1497 0.16
Bagging(RandF) 91.9469 0.994 0.9059 0.1381 2.56
2. ReliefF:11
FURIA 94.5523 0.985 0.9364 0.1142 1.02
C4.5 94.3155 0.980 0.9336 0.1231 0.01
RandF 95.5471 0.998 0.948 0.1098 0.28
Bagging(FURIA) 96.684 0.998 0.9613 0.094 8.32
Bagging(C4.5) 95.5945 0.997 0.9485 0.0981 0.11
Bagging(RandF) 95.5945 0.998 0.9485 0.1167 2.42
3. Pairwise Consistency:9
FURIA 95.3103 0.988 0.9452 0.1075 1.12
C4.5 94.5997 0.980 0.9369 0.1203 0.03
RandF 96.7788 0.999 0.9624 0.0998 0.48
Bagging(FURIA) 96.7314 0.998 0.9618 0.096 10.36
Bagging(C4.5) 95.5945 0.998 0.9485 0.0974 0.13
Bagging(RandF) 94.0313 0.973 0.9303 0.1284 0.17
4. Pairwise Correlation:8
FURIA 86.8783 0.955 0.8467 0.1748 1.19
C4.5 86.5467 0.948 0.8428 0.1865 0.02
RandF 92.0417 0.994 0.907 0.1345 0.36
Bagging(FURIA) 90.3837 0.990 0.8877 0.1436 13.84
Bagging(C4.5) 88.9626 0.985 0.8711 0.1531 0.16
Bagging(RandF) 91.9469 0.994 0.9059 0.1402 3.18
5. The Inconsistency Metric :7

49
FURIA 94.8839 0.986 0.9402 0.1094 1.03
C4.5 94.3155 0.978 0.9336 0.124 0.01
RandF 96.5419 0.999 0.9596 0.1033 0.3
Bagging(FURIA) 96.5419 0.998 0.9596 0.0934 9.4
Bagging(C4.5) 95.7366 0.997 0.9502 0.0988 0.11
Bagging(RandF) 96.3998 0.999 0.9579 0.1106 2.65
6. TabuSearch:7
FURIA 94.3629 0.984 0.9342 0.1139 1.59
C4.5 94.2681 0.980 0.9331 0.1229 0.08
RandF 95.9261 0.998 0.9524 0.1092 0.43
Bagging(FURIA) 96.5419 0.998 0.9596 0.0964 9.82
Bagging(C4.5) 95.3103 0.998 0.9452 0.1004 0.13
Bagging(RandF) 95.784 0.998 0.9508 0.1168 2.63
Table 7 reports the performance after altering the original dataset. The RandF classifier
(97.6788%) outperformed the FURIA (94.5997%) and C4.5 (93.9839%) classifiers. At the same
time, the Bagging based on the RandF classifier outperformed the other classifiers by scoring
accuracy of 97.5367% while using 19 features.
Table 8 lists the obtained features using mRMR, ReliefF, Pairwise Consistency, Pairwise
Correlation, The Inconsistency Metric, and Tabu search. This step has reduced the feature size
from 19 to 11 and 6 features only. Figure ?? shows the resulting Venn diagram, which
demonstrates the relationship among Feature selection Techniques. Table 9 illustrates all the
possible intersections between Feature selection Techniques using the original dataset. Worth
noting that all the feature selection techniques share Weight and BFP attributes. Most methods
share Age, Gender, FHWO, FCVC, and CAEC features.
Table 10 reports the classification performance after performing feature selection over the
modified dataset. The RandF classifier outperformed the other classifiers achieving an accuracy
of 97.63% using the Pairwise Consistency that yielded 15 features. At the same time, the Bagging
based on the RandF algorithm outperformed the other classifiers by scoring an accuracy of
97.58% while using the Pairwise Consistency technique. As the feature selection methods
reduced the features of datasets, it also increased the overall performance accuracy, as with the
Pairwise Consistency and TabuSearch techniques.
Although the number of features chosen by the Pairwise Consistency, Pairwise Correlation, and
Inconsistency Metric techniques is less than the initial dataset, the classification results were
better. The Inconsistency Metric, Pairwise Consistency, and Pairwise Correlation were the best
feature selection techniques. The Bagging algorithm enhanced the prediction accuracy. Overall,
the performance of the Bagging algorithm is more promising than that of the base algorithms
independently. It is worth noting that the construction time for the model is reduced, especially
for The Inconsistency Metric, Pairwise Consistency, and Pairwise Correlation techniques. The
recent results indicate that the feature reduction stage successfully helps in improving
classification accuracy. As the filter-based feature selection method reduces the features of
datasets, it also decreases the time taken to build the model and increases the overall
performance. Based on the results of experiments, Pairwise Consistency and Pairwise Correlation
techniques are shown to be promising tools for feature selection in respect of the quality of
obtained feature subset and computation efficiency.

50
Table 8. Classification performance without feature selection methods for the modified dataset
Classifier ACC(%) AUC Kappa RMSE Time(s)
FURIA 94.5997 0.984 0.9369 0.1106 1.58
C4.5 93.9839 0.976 0.9297 0.1288 0.11
RandF 97.6788 0.999 0.9729 0.088 0.45
Bagging(FURIA) 97.2525 0.998 0.9679 0.086 8.7
Bagging(C4.5) 95.0261 0.995 0.9419 0.1066 0.19
Bagging(RandF) 97.5367 0.999 0.9712 0.0933 2.9
Table 9. Optimal feature subsets the modified dataset
Feature Selection Method No. of
Features
Features
1. mRMR 11 Gender, Age, Weight, FHWO, FAVC, FCVC,
NCP, CAEC, CH2O, TUE, BFP
2. ReliefF 11 BFP, Gender, Weight, RMR, BMR, FHWO,
FCVC, CALC, CAEC, Height, FAF
3. Pairwise Consistency 15 BFP, Weight, BMR, RMR, Age, FCVC, FAF,
Height, CH2O, TUE, NCP, Gender, CAEC,
CALC, FHWO
4. Pairwise Correlation 14 BFP, Weight, RMR, BMR, FCVC, Age,
TUE, NCP, Gender, CH2O, FAF, FHWO,
CAEC, Height
5. The Inconsistency Metric 6 Age, Height, Weight, FAF, CALC, BFP
6.Tabu 11 Gender, Age, Weight, FHWO, FAVC, FCVC,
NCP, CAEC, CH2O, TUE, BFP
Figure 2. Venn diagram demonstrating the cross-links between the feature selection techniques used over
the modified dataset
Table 10. Intersecting Features among Feature selection Techniques using the modified dataset
Feature Selection Techniques Intersecting Features
[mRMR ] and [ReliefF] and [PairwiseCons.] and [PairwiseCorr.]
and [Inconsistency.M] and [TabuSearch]:
Weight
BFP
Age

51
[mRMR ] and [ReliefF] and [PairwiseCons.] and [PairwiseCorr.]
and [TabuSearch]:
Gender
FHWO
FCVC
CAEC
[TabuSearch]:
NCP
CH2O
TUE
[ReliefF] and [PairwiseCons.] and [Inconsistency.M] CALC
[ReliefF] and [PairwiseCons.] and [PairwiseCorr.]: RMR
BMR
[ReliefF] and [PairwiseCons.] and [PairwiseCorr.] and
[Inconsistency.M]
Height
FAF
[mRMR ] and [TabuSearch] FAVC
Table 11. Classification performance after feature selection for the modified dataset
1. mRMR:11
FURIA 94.6471 0.986 0.9375 0.1095 1.41
C4.5 93.7944 0.975 0.9275 0.1281 0.08
RandF 97.2051 0.999 0.9674 0.0937 0.45
Bagging(FURIA) 96.3051 0.998 0.9568 0.0907 7.09
Bagging(C4.5) 94.3629 0.995 0.9342 0.1084 0.13
Bagging(RandF) 96.8261 0.999 0.9629 0.0988 2.38
2. ReliefF:11
FURIA 94.8839 0.984 0.9402 0.1139 0.86
C4.5 93.747 0.974 0.927 0.1307 0.02
RandF 97.063 0.999 0.9657 0.0872 0.26
Bagging(FURIA) 96.0682 0.997 0.9541 0.0947 7.26
Bagging(C4.5) 95.1208 0.997 0.943 0.1043 0.13
Bagging(RandF) 96.8735 0.999 0.9635 0.0919 2.35
3. Pairwise Consistency:15
FURIA 95.0261 0.985 0.9419 0.1083 0.98
C4.5 94.2207 0.975 0.9325 0.1261 0.02
RandF 97.6315 0.999 0.9723 0.0888 0.29
Bagging(FURIA) 96.684 0.999 0.9613 0.0878 8.24
Bagging(C4.5) 95.2629 0.997 0.9447 0.1042 0.18
Bagging(RandF) 97.5841 0.999 0.9718 0.0939 2.69
4. Pairwise Correlation:14
FURIA 94.8839 0.986 0.9402 0.1078 1.00
C4.5 94.126 0.975 0.9314 0.1269 0.02
RandF 97.4893 0.999 0.9707 0.0875 0.31
Bagging(FURIA) 96.5893 0.999 0.9602 0.0881 7.91
Bagging(C4.5) 95.1208 0.996 0.943 0.1047 0.18
Bagging(RandF) 97.5367 0.999 0.9712 0.0934 2.78
5. The Inconsistency Metric :6
FURIA 96.0208 0.987 0.9535 0.1001 0.7
C4.5 94.5523 0.976 0.9364 0.1221 0.01

52
RandF 97.5367 0.999 0.9712 0.0792 0.26
Bagging(FURIA) 97.0156 .998 0.9651 0.0862 6.69
Bagging(C4.5) 95.1208 0.997 0.943 0.1015 0.1
Bagging(RandF) 97.2999 0.999 0.9685 0.0854 2.27
6. TabuSearch:11
FURIA 94.6471 0.986 0.9375 0.1095 0.95
C4.5 93.7944 0.975 0.9275 0.1281 0.02
RandF 97.2051 0.999 0.9674 0.0937 0.27
Bagging(FURIA) 96.3051 0.998 0.9568 0.0907 7.24
Bagging(C4.5) 94.3629 0.995 0.9342 0.1084 0.13
Bagging(RandF) 96.8261 0.999 0.9629 0.0988 2.43
Table 12. Summary of intersecting features among Inconsistency Metric, Pairwise Consistency, and
Pairwise Correlation feature selection techniques
Dataset Intersecting Features
Original dataset Weight
Age
FAF
CH2O
TUE
Modified dataset BFP
Weight
Age
FAF
Height
[original] and [modified] Weight
Age
FAF
The analysis of the results obtained from the original and modified datasets has established a
relationship between obesity/overweight and common risk factors such as weight, age, and
physical activity patterns. The results from the original dataset show that the techniques
mentioned above have weight, age, FAF, CH2O, and TUE features in common. While in the
modified dataset, the techniques share the weight, age, Height, FAF, and BFP features. Table 10
lists intersecting features among Inconsistency Metric, Pairwise Consistency, and Pairwise
Correlation feature selection techniques. In sum, our findings are consistent with the findings
from studies of [13,15,16,17,18,19] mentioned earlier in Section 2.
CONCLUSION
Obesity exists; this article seeks to explore the risk factors behind this disease from a data-mining
point of view. The primary objective is to improve the prediction accuracy of obesity with a
minimal number of feature subsets using the dataset gathered previously. For that reason, we
adapted a bagging algorithm based on filter-based feature selection to improve the prediction
accuracy of obesity with a minimal number of feature subsets. We utilized several machine
learning algorithms for classifying the obesity classes and several filter feature selection methods
to maximize the classifier accuracy. The proposed work improves the classification accuracy
compared to the previous work from the experimental result. Experiments on the original and
modified dataset proved that our proposed method could reduce the number of features by almost
97% and obtain satisfactory results.

53
REFERENCES
[1] WHO Page. Available online: https://www.who.int/news-room/fact-sheets/detail/obesity-and-
overweight; https://www.who.int/nmh/publications/ncd_report_chapter1.pdf (accessed on 18 March
2020).
[2] Ward, Z.J.; Bleich, S.N.; Cradock, A.L.; Barrett, J.L.; Giles, C.M.; Flax, C.; Long, M.W.; Gortmaker,
S.L. Projected US State-Level Prevalence of Adult Obesity and Severe Obesity. N. Engl. J. Med.
2019, 381, 2440–2450.
[3] Csige, I.; Ujvárosy, D.; Szabó, Z.; Lo ̋rincz, I.; Paragh, G.; Harangi, M.; Somodi, S. The impact of
obesity on the cardiovascular system. J. Diabetes Res. 2018, 2018, 3407306.
[4] Braet, C., 2007. Childhood obesity. Eating disorders in children and adolescents, pp.182-192.
[5] Lindberg, L., Hagman, E., Danielsson, P., Marcus, C. and Persson, M., 2020. Anxiety and depression
in children and adolescents with obesity: a nationwide study in Sweden. BMC medicine, 18(1), pp.1-
9.
[6] Oppert, J.M., Bellicha, A., van Baak, M.A., Battista, F., Beaulieu, K., Blundell, J.E., Carraça, E.V.,
Encantado, J., Ermolao, A., Pramono, A. and Farpour‐Lambert, N., 2021. Exercise training in the
management of overweight and obesity in adults: Synthesis of the evidence and recommendations
from the European Association for the Study of Obesity Physical Activity Working Group. Obesity
Reviews, 22, p.e13273.
[7] Dogan, A. and Birant, D., 2021. Machine learning and data mining in manufacturing. Expert Systems
with Applications, 166, p.114060.
[8] dos Santos, B.S., Steiner, M.T.A., Fenerich, A.T. and Lima, R.H.P., 2019. Data mining and machine
learning techniques applied to public health problems: A bibliometric analysis from 2009 to 2018.
Computers & Industrial Engineering, 138, p.106120.
[9] dos Santos, B.S., Steiner, M.T.A., Fenerich, A.T. and Lima, R.H.P., 2019. Data mining and machine
learning techniques applied to public health problems: A bibliometric analysis from 2009 to 2018.
Computers & Industrial Engineering, 138, p.106120.
[10] Li, Y., Li, T. and Liu, H., 2017. Recent advances in feature selection and its applications. Knowledge
and Information Systems, 53(3), pp.551-577.
[11] Singh, B.; Tawfik, H. A Machine Learning Approach for Predicting Weight Gain Risks in Young
Adults. In Proceedings of the 10th IEEE International Conference on Dependable Systems, Services
and Technologies (DESSERT), Leeds, UK, 5–7 June 2019; pp. 231–234.
[12] Singh, B. and Tawfik, H., 2020, June. Machine learning approach for the early prediction of the risk
of overweight and obesity in young people. In International Conference on Computational Science
(pp. 523-535). Springer, Cham.
[13] Farran, B.; AlWotayan, R.; Alkandari, H.; Al-Abdulrazzaq, D.; Channanath, A.; Thangavel, A.T. Use
of Non-invasive Parameters and Machine-Learning Algorithms for Predicting Future Risk of Type 2
Diabetes: A Retrospective Cohort Study of Health Data From Kuwait. Front. Endocrinol. 2019, 10,
624.
[14] Padmanabhan, M.; Yuan, P.; Chada, G.; van Nguyen, H. Physician-friendly machine learning: A case
study with cardiovascular disease risk prediction. J. Clin. Med. 2019, 8, 1050.
[15] Jindal, K.; Baliyan, N.; Rana, P.S. Obesity Prediction Using Ensemble Machine Learning
Approaches. In Recent Findings in Intelligent Computing Techniques; Springer: Singapore, 2018; pp.
355–362.
[16] Selya, A.S.; Anshutz, D. Machine Learning for the Classification of Obesity from Dietary and
Physical Activity Patterns. In Advanced Data Analytics in Health; Springer: Cham, Switzerland,
2018; pp. 77–97.
[17] Dunstan, J.; Aguirre, M.; Bastías, M.; Nau, C.; Glass, T.A.; Tobar, F. Predicting nationwide obesity
from food sales using machine learning. Health Inform. J. 2019. Available online:
https://journals.sagepub.com/doi/full/ 10.1177/1460458219845959 (accessed on 18 March 2020).
[18] Zheng, Z.; Ruggiero, K. Using machine learning to predict obesity in high school students. In
Proceedings of the 2017 IEEE International Conference on Bioinformatics and Biomedicine (BIBM),
Kansas City, MO, USA, 13–16 November 2017; pp. 2132–2138.
[19] DeGregory, K.W.; Kuiper, P.; DeSilvio, T.; Pleuss, J.D.; Miller, R.; Roginski, J.W.; Fisher, C.B.;
Harness, D.; Viswanath, S.; Heymsfield, S.B.; et al. A review of machine learning in obesity. Obes.
Rev. 2018, 19, 668–685.

54
[20] Golino, H.F.; Amaral, L.S.D.B.; Duarte, S.F.P.; Gomes, C.M.A.; Soares, T.D.J.; Reis, L.A.D.;
Santos, J. Predicting increased blood pressure using machine learning. J. Obes. 2014, 2014, 637635.
[21] Pleuss, J.D.; Talty, K.; Morse, S.; Kuiper, P.; Scioletti, M.; Heymsfield, S.B.; Thomas, D.M. A
machine learning approach relating 3D body scans to body composition in humans. Eur. J. Clin. Nutr.
2019, 73, 200–208.
[22] Maharana, A.; Nsoesie, E.O. Use of deep learning to examine the association of the built environment
with prevalence of n eighborhood adult obesity. JAMA Netw. Open 2018, 1, e181535.
[23] Pouladzadeh, P.; Kuhad, P.; Peddi, S.V.B.; Yassine, A.; Shirmohammadi, S. Food calorie
measurement using deep learning neural network. In Proceedings of the 2016 IEEE International
Instrumentation and Measurement Technology Conference, Taipei, Taiwan, 23–26 May 2016; pp. 1–
6.
[24] De-La-Hoz-Correa, E., Mendoza Palechor, F., De-La-Hoz-Manotas, A., Morales Ortega, R. and
Sánchez Hernández, A.B., 2019. Obesity level estimation software based on decision trees. J Comput
Sci, 15 (Issue 1) (2019), pp. 67-77, 10.3844/jcssp.2019.67.77 .
[25] Hsu, P.H., Lee, C.H., Kuo, L.K., Kung, Y.C., Chen, W.J. and Tzeng, M.S., 2018. Determination of
the energy requirements in mechanically ventilated critically ill elderly patients in different BMI
groups using the Harris–Benedict equation. Journal of the Formosan Medical Association, 117(4),
pp.301-307.
[26] Wu, G., 2016. Dietary protein intake and human health. Food & function, 7(3), pp.1251-1265.
[27] Deurenberg, P., Weststrate, J.A. and Seidell, J.C., 1991. Body mass index as a measure of body
fatness: age-and sex-specific prediction formulas. British journal of nutrition, 65(2), pp.105-114.
[28] Wan, C., 2019. Feature selection paradigms. In Hierarchical Feature Selection for Knowledge
Discovery (pp. 17-23). Springer, Cham.
[29] Lee, S.J., Xu, Z., Li, T. and Yang, Y., 2018. A novel bagging C4. 5 algorithm based on wrapper
feature selection for supporting wise clinical decision making. Journal of biomedical informatics, 78,
pp.144-155.
[30] Hedar, A.R., Wang, J. and Fukushima, M., 2008. Tabu search for attribute reduction in rough set
theory. Soft Computing, 12(9), pp.909-918.
[31] Prajapati, V.K., Jain, M. and Chouhan, L., 2020, February. Tabu search algorithm (TSA): A
comprehensive survey. In 2020 3rd International Conference on Emerging Technologies in Computer
Engineering: Machine Learning and Internet of Things (ICETCE) (pp. 1-8). IEEE.
[32] Peng, H., Long, F. and Ding, C., 2005. Feature selection based on mutual information criteria of max-
dependency, max-relevance, and min-redundancy. IEEE Transactions on pattern analysis and
machine intelligence, 27(8), pp.1226-1238
[33] Urbanowicz, R.J., Meeker, M., La Cava, W., Olson, R.S. and Moore, J.H., 2018. Relief-based feature
selection: Introduction and review. Journal of biomedical informatics, 85, pp.189-203.
[34] Liu, J.C., 1996. Optimality and duality for generalized fractional programming involving nonsmooth
(F, ρ)-convex functions. Computers & Mathematics with Applications, 32(2), pp.91-102.
[35] Jiménez, F., Palma, G.S.J., Miralles-Pechuán, L. and Botía, J., 2021. Multivariate feature ranking of
gene expression data. arXiv preprint arXiv:2111.02357.
[36] Liu, H. and Setiono, R., 1996, July. A probabilistic approach to feature selection-a filter solution. In
ICML (Vol. 96, pp. 319-327).
[37] Quinlan, J.R., 2014. C4. 5: programs for machine learning. Elsevier.
[38] Breiman, L., 2001. Random forests. Machine learning, 45(1), pp.5-32.
[39] Hühn, J. and Hüllermeier, E., 2009. FURIA: an algorithm for unordered fuzzy rule induction. Data
Mining and Knowledge Discovery, 19(3), pp.293-319.
[40] Breiman, L., 1996. Bagging predictors. Machine learning, 24(2), pp.123-140.
[41] Lee, S.J., Xu, Z., Li, T. and Yang, Y., 2018. A novel bagging C4.5 algorithm based on wrapper
feature selection for supporting wise clinical decision making. Journal of biomedical informatics, 78,
pp.144-155.
[42] Fernández, A., García, S., Galar, M., Prati, R.C., Krawczyk, B. and Herrera, F., 2018. Performance
measures. In Learning from Imbalanced Data Sets (pp. 47-61). Springer, Cham.

DIAGNOSIS OF OBESITY LEVEL BASED ON BAGGING ENSEMBLE CLASSIFIER AND FEATURE SELECTION METHODS

More Related Content

Similar to DIAGNOSIS OF OBESITY LEVEL BASED ON BAGGING ENSEMBLE CLASSIFIER AND FEATURE SELECTION METHODS (20)

Recently uploaded (20)

DIAGNOSIS OF OBESITY LEVEL BASED ON BAGGING ENSEMBLE CLASSIFIER AND FEATURE SELECTION METHODS