IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 1, February 2025, pp. 376∼384

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp376-384 ❒ 376

A systematic analysis on machine learning classifiers with

data pre-processing to detect anti-pattern from source code
Nazneen Akhter1 , Afrina Khatun1 , Md. Sazzadur Rahman2 , A. S. M. Sanwar Hosen3 ,
Mohammad Shahidul Islam2
1 Departmentof Computer Science and Engineering, Bangladesh University of Professionals, Dhaka, Bangladesh
2 Institute of Information Technology, Jahangirnagar University, Dhaka, Bangladesh
3 Department of Artificial Intelligence and Big Data, Woosong University, Daejeon, South Korea

Article Info ABSTRACT

Received Jan 16, 2024 Automatic detection of anti-patterns from source code can reduce software
maintenance costs massively. Nowadays, machine learning approaches are very
Revised Jul 10, 2024
commonly used to identify anti-patterns. Hence, it is very crucial to choose a
Accepted Jul 26, 2024
classifier that can be useful for detecting anti-patterns. This work aims to help
practitioners to choose a suitable classifier to detect anti-patterns. In this paper,
Keywords: we highlight 16 classifiers in four different categories to detect anti-patterns.
Furthermore, the performance of these classifiers is identified with the data
Anti-pattern
pre-processing (DPP) to detect four commonly occurring anti-patterns from the
Code quality three commonly used open-source Java projects’ source code. The accuracy of
Code smell Dagging classifiers is 98.4%. Kernel logistic regression (KLR) also performs
Data pre-processing well i.e., 97%. In the case of time complexity, naive Bayes (NB), decision
Machine learning trees (DT), support vector machines (SVM), library for support vector machines
(LibSVM), logistic, and LightGBM (LB) have less time complexity to build a
model in all the projects.

This is an open access article under the CC BY-SA license.

Corresponding Author:
A. S. M. Sanwar Hosen
Department of Artificial Intelligence and Big Data, Woosong University
Daejeon 34606, South Korea
Email: sanwar@wsu.ac.kr

1. INTRODUCTION
Software quality is very important as it defines the ease of maintainability, testability, readability,
stability, speed, usability, size, cost, and security. It also ensures the smoothness, conciseness, and customer
satisfaction of software. However, due to the heavy workload and time pressure, software quality issues are
usually ignored. Some reports show that software evolution, understandability, usability, modularity, reusabil-
ity, analyzability, and changeability are neglected during the development process of the software [1]. As a
result, design flaws i.e., poor structure in the source code increase the development and maintenance cost.
These design flaws are termed as anti-pattern or code smell. Palomba et al. [2] report that the maintenance
cost of software is 2–100 times greater than the development cost. The term code smell (anti-pattern) is first
introduced in [3]. Anti-patterns from a source code can be detected both by manual and automated processes.
For large software projects, the manual detection system is a very time-consuming process as it mainly de-
pends on proper documentation, source code structure, and developer’s experience [1]. Hence, an automated
and effective anti-pattern detection technique is essential. Although few studies are available in this field.
Sometimes people may become confused about choosing machine learning (ML) classifiers and environmental

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938 ❒ 377

setup. Several researchers introduced three different support vector machines (SVM)-based approaches [4]–[6].
Maiga et al. did not adopt any data pre-processing (DPP) technique, while Akhter et al. adopted a DPP
technique namely synthetic minority over-sampling technique (SMOTE) with SVM which performed better
than synthetic minority under-sampling with random forest (SMURF). Five different ML classifiers: J48,
random forest (RF), naive Bayes (NB), SVM, and just-in-time rule induction procedure (JRIP) to detect anti-
pattern are used in [7]. They reported NB performed the best. They did a comparison among ML classifiers and
detection of code and design smells (DECOR) (a method of code and design smell detection). They reported
that in terms of F-measure and recall, the performance of DECOR was higher than NB but precision was lower.
The author used 16 different classifiers and their boosting techniques (in total 32 classifiers) [8]. They reported
that the best algorithm had approximately 97% of F-Score.
The aforementioned researchers hardly use any techniques to process the data. In comparison, a
DPP technique can influence the performance of the classifier. As there are many classifiers most of the time
researchers become puzzled to choose a classifier. Hence, an appropriate classifier selection is also a crucial
study for anti-pattern detection. In this paper, a comprehensive study is conducted so that the practitioners can
get a complete idea of choosing the classifier and their working process. A data-prepossessing technique is
adopted to enhance the performance of the classifiers.
This is how the remainder of the paper is structured. The background research on anti-pattern detection
is covered in section 2. In section 3, an explanation of the suggested methodology is provided along with an
analysis of sixteen ML classifiers. Section 4 addresses the outcome and execution. Section 5 concludes this
report by discussing future research.

2. LITERATURE REVIEW ON ANTI-PATTERN DETECTION

Inadequate design raises the system’s development and maintenance costs by lowering performance
and reusability. As a result, scholars developed various theories to address this issue. The three main types of
methodologies used in anti-pattern detection research that are currently in use are rule-based, ML-based, and
deep learning-based. Detailed discussion is provided in the following.
The main aim of the rule-based approach is to propose a set of rules to find anti-patterns. To do
that, several researchers [9], [10] propose two different approaches to identify anti-patterns from the source
code. Blob, spaghetti code, and functional decomposition by utilizing three heuristic search algorithms [9],
[10]. Aras and Selcuk [11] present an automated approach using metrics and rules to detect anti-patterns in an
object-oriented system. Three anti-patterns are used namely Blob, Swiss Army Knife, and Lava Flow. 36 Java
classes are focused on conducting this experiment. However, the accuracy is not satisfactory. Due to issues with
rule-based systems’ accuracy, researchers introduced ML techniques. The primary issue with the rule-based
technique is that each smell has its own set of rules, all of which are manually specified by the researcher.
A method for identifying design patterns utilizing ML-based techniques and software metrics has been
put out by [12]. For small scale, code recall is 90% but for large scale code, it is 60%. Two SVM-based methods,
SMURF and SVMDetect, are introduced by Maiga et al. [5], [6] to identify four anti-patterns: Swiss Army
Knife, Blob, functional decomposition, and spaghetti code. The precision and recall for SMURF are 97%
and 84.09% respectively (Blob Class). After computation, they report that SVMDetect can detect 143 anti-
patterns whereas DETEX only detects 102. The recall and precision rates of SVMDetect are 84.09 and 97.09%
respectively for (Blob class). In terms of anti-pattern identification, both methods perform similarly and can
identify 143 instances of blobs within the system. Barbez et al. [13] present an ensemble method called smart
aggregation of anti-patterns detectors (SMAD) to detect god class and feature envy. SMAD combined different
approaches depending on their internal detection rule. Yin et al. [14] reported that accuracy is more than
85% for the classifier AB-J48-pruned, RF, and AdaBoost random forest (AB-RF) among 15 ML classifiers.
Azadi et al. [15] propose a tool named WekaNose to investigate code smell in a system. They intend to
characterize the instances that influence the detection of code smell or not. Cruz et al. [16] introduce a detection
technique with seven different machine-learning algorithms for detecting four types of bad smells.
Kumar and Sureka [17] develop a predictive model to detect anti-patterns. They have mentioned that
RF performed the best. Pritam et al. [18] examine the impact of code smell on the change tendency of a specific
class in an item system which leads them to discover mistakes in expectation of progress inclination utilizing
code smell. They use 4100 unique classes of 14 software systems after per-processing. They have reported
the sensitivity is 0.70 and specificity higher than 0.67 for multi-layer perceptron (MLP). Jesudoss et al. [19]

A systematic analysis on machine learning classifiers with data pre-processing to ... (Nazneen Akhter)
378 ❒ ISSN: 2252-8938

propose another code smell detection technique utilizing two algorithms such as SVM and RF. SVM acted as a
classifier and the RF is used for predicting the range of data. From the above discussion, we hardly found any
DPP adopted. As a result, the accuracy rate of the classifiers decreases. Most of the researchers used the same
type of ML classifiers like SVM, RF, and NB. It is a matter of concern that other classifiers like ensemble or
meta classifiers performance are missing to detect anti-pattern.
Kacem et al. [20] propose a hybrid learning approach to identify four code smells namely god
class, data class, feature envy, and long method from 74 open-source systems. They have reported that the
performance of god class was very good (god class precision 99.28%, recall 98.58%, F-measure 98.93%).
Das et al. [21] come up with a technique named convolution neural networks (CNN) to detect code smells.
They have achieved 97% accuracy for the Brain class and 95% for the Brain method. Barbez et al. [22] propose
a method namely convolutional analysis of code metrics evolution (CAME) to detect anti-patterns. From the
background study, it is found that DPP techniques are missing. ML classifiers can enhance their performance
if the model gets pre-processed data. Recent research works are using similar kinds of ML classifiers (the
most used classifiers are SVM, NB, and RF). Sometimes it becomes very tough to choose a classifier for the
practitioner to work in this field.

3. ANTI-PATTERN DETECTION MODEL

In this paper, a constructive analysis of ML classifiers is presented to detect anti-patterns. Sixteen
different types of classifiers have been adopted to conduct the study. Considering their nature, classifiers are
divided into four subgroups: function classifiers, Bayes classifiers, tree classifiers, and meta classifiers [23].
Figure 1 represents the details of the proposed model. A detailed description of those is provided to understand
the working principle of the classifiers deeply in the case of anti-pattern detection. The following describes the
specifics.

Figure 1. The proposed methodology to detect anti-pattern from source code

3.1. Dataset and pre-processing

Three separate open-source Java projects—Azureus v2.3.0.6, Xerces v2.7.0, and ArgoUML v0.19.8
are employed in this study. ArgoUML is an application for creating unified modeling language (UML) dia-
grams. The ArgoUML dataset comprises 1,230 classes. Azureus with 1,449 classes uses the bit-torrent protocol
to transmit data. The XML software library is called Xerces. There are 513 classes in the Xereces data file.
This paper uses data sets supplied in [6]. To balance the data and improve the characteristics, we examined
SMOTE as a DPP technique in this paper. This technique randomly chooses the minority class instance, and
the k closest minority class neighbors are found. Afterward, a line segment in the feature space is formed by
joining a and b, one of the k nearest neighbors (b), at random to generate the synthetic instance. The two
selected examples, a and b, are used to create the synthetic instances by using the convex combination [24].
Consequently, new synthetic instances are generated that can be referred to as feature-boosting techniques.

3.2. Critical analysis of classifiers on anti-pattern detection

Functional classifiers mainly include the idea of neural networks and regression. The main procedure
of a functional classifier is to assign data labels based on a function that is also known as the core function
of the classifier. In this work, five functional classifiers are considered. The first classifier SVM performs the
classification by calculating a boundary value known as a hyperplane that differentiates two data classes. SVM
is broadly categorized into two groups; linear and non-linear SVM [25]. Non-linear SVM utilizes kernel tricks
to classify data. The kernel function is used to transform lower dimensional input space into higher dimensional
space. Kernel logistic regression (KLR) solves the classification problems by transforming the real input space
into a high-dimensional feature space by utilizing kernel functions [26]. LibSVM is a complete package of
SVM. This classifier can support various SVM formulations for classification [27]. In this project, support

Int J Artif Intell, Vol. 14, No. 1, February 2025: 376–384

Int J Artif Intell ISSN: 2252-8938 ❒ 379

vector classification (SVC) (two-class and multi-class) is used. Logistic regression predicts the odds ratio
which is the event that needs to be predicted, for example, how likely a sample is an anti-pattern [28]. Multi-
layer perceptron classifier (MLPCL) can train fully connected MLP networks with one hidden layer. For further
processing, all attributes are standardized (it gives all features the same influence). In this project, conjugate
gradient descent is used as it can reduce the time complexity.
The Bayes classifier provides a probabilistic model. that makes the most probable prediction for a new
example. The Bayes theorem is used to calculate the conditional probability. Naive Bayes (NB) is a classifier
that uses a probabilistic approach for the classification tasks. NB follows the following equation to calculate the
posterior probability for each class [29]. The classes that have the highest posterior probability are the output
of prediction.
Tree classifiers divide the training data into subsets by utilizing a sequence of conditional statements.
Naive-Bayesian tree (NBTree) classifier combines a decision tree (DT) with a NB classifier [30]. The tree is
constructed recursively to represent the attributes of the dataset. The nodes of the tree utilize the DT classifier.
On the other hand, the leaves contain naive-Bayesian classifiers [31]. DT utilizes instances of data to create
trees. When the tree construction is done, it prunes the trees to improve the model’s performance. In this work,
an advanced decision tree algorithm namely the C4.5 algorithm is used (called J48 in Weka) [32]. Decision
stump (DS) is simply a DT that utilizes a single attribute for splitting [33]. It creates a single root tree of the
given values using a top-down greedy approach. RF creates a forest of randomly generated DT. RF combines
the output of the DT to form the final output. The number of trees defines the accuracy of the classifiers. For
example, the more trees, the more accuracy is produced. In our implementation, 100 trees are created [34].
Meta classifiers can ensemble different classifiers to enhance the performance of the base classifiers.
Logit boost (LB) is a meta classifier that is based on the AdaBoost classifier. Among the weighted sample,
higher weights will be assigned to the unclassified data [35]. Real AdaBoost (RAB) is a boosting classifier
for the binary classification task. Here, the weak learner algorithm provides class probability estimation [36].
Probability estimation:

P (x) = P w(Y = 1|x)ϵ[0, 1] (1)

Here, w is the weight of the training data. In this study, the DS is used as the base classifier. The number of
iterations is considered 10. It updates the weights on every iteration.
Bagging is an ensemble learner that randomly redistributes the original training dataset for individual
base models [37]. Voting or averaging their prediction probabilities gives a final prediction. Diverse ensemble
creation by oppositional re-labeling of artificial training examples (DECORATE) is designed to use additional
artificially generated training data [37]. Additionally, it achieves equivalent performance on bigger training
sets and gains higher accuracy than boosting on the small training set [38]. Dagging creates several disjoint,
stratified folds of the data to predict the final output by averaging them [39]. Voting is another ensemble
algorithm that creates two or more sub-models with the base classifiers to classify data. The final prediction is
done by aggregating the sub-model predictions. In majority voting, the final prediction is made by combining
the sub-models output. The class that gets the most number of votes will be the output of the classification [40].

4. RESULT ANALYSIS
The performance analysis of various ML classifiers is presented in this section. Five evaluation metrics
are used: accuracy, F-measure, precision, recall, and receiver operating characteristic (ROC). An Intel Core i3
processor with 4 GB RAM and a 64-bit operating system is used for the experiment. WEKA is employed in
the implementation of the classifier models. The details are described in the following.
4.1. Research questions and evaluation
Three research issues are addressed in this section to validate the anti-pattern detection technique. In
RQ1, the significance of the DPP technique is briefly discussed. The performance of classifiers in terms of
computing time is shown in RQ2. Each study question is thoroughly described and evaluated in the analysis
that follows.
4.1.1. RQ1: How well do classifiers work with data pre-processing techniques to find anti-patterns?
In the project of ArgoUML, the Dagging classifier performs better in terms of accuracy than other
classifiers. The accuracy of the Dagging is 98.4% when DS is used as a base classifier. The accuracy of DS as

A systematic analysis on machine learning classifiers with data pre-processing to ... (Nazneen Akhter)
380 ❒ ISSN: 2252-8938

an individual classifier is 91.37%. Dagging breaks the original training dataset into small chunks. With these
small chunks of data, the base classifier can train itself well. So, it can classify data more accurately from the
test dataset. For this dataset, the NB classifier performs poorly in terms of accuracy, 79.37%. This is because
of the high dependency on the dataset attributes. The Dagging classifier is also used on the dataset using the
NB as the base classifier. For that case, the accuracy of Dagging becomes 91.99%. Hence, it can be said that
meta-classifiers like Dagging can enhance the performance of weak-performed classifiers. Table 1 shows the
detailed result of the classifiers in terms of accuracy for ArgoUML, Azureus, and Xerces project. In the project
of Azureus, data are non-linearly separable. The accuracy of KLR is 97%. This is because KLR can perform
well with non-linearly separate data. KLR offers a natural estimate of the class probability. On the other side,
the Dagging classifier performed poorly for this project. DS is used as the base classifier.

Table 1. Accuracy of the classifiers for ArgoUML, Azureus, and Xerces projects
Classifiers Accuracy (ArgoUML) (%) Accuracy (Azureus) (%) Accuracy (Xerces) (%)
SVM 92.48 90.32 91.49
NB 79.37 76.51 73.71
RF 92.78 92.03 93.35
DT 90.3 91.15 87.76
LibSVM 96.55 90.37 95.82
Logistic 90.78 81.27 88.58
KLR 91.82 96.99 89.90
NBTree 88.03 85.31 86.60
MLPCL 91.86 93.93 90.285
RAB 92.31 92.61 93.08
Bagging 91.78 91.8 88.47
Dagging 98.37 68.69 96.04
LB 91.82 96.01 92.59
DECORATE 92.92 91.67 92.75
DS 91.37 87.26 89.02

Performance of tree classifiers: in this work, four tree classifiers have been used; DT, RF, NBTree, and
DS. Among those, the RF classifier performs better than other classifiers for all three projects. The accuracy
of RF for ArgoUML, Azureus, and Xerces are 92.8%, 92.03%, 93.4%, respectively. This is because the RF
is an ensemble learner that finds the best features from the dataset randomly. The final prediction is done by
combining the output of many decision trees. NBTree performs lower than other tree classifiers. The accuracy
of NBTree for the three projects is 88.03%, 84.81%, and 86.61%, respectively. NBTree utilizes NB in the
leaves of the tree. NB tree shows minimum performance for the dependent dataset which is the main reason
behind performing low for NBTree. Performance of ensemble classifiers: for the project of ArgoUML, Dagging
classifier performs the best whose accuracy is 98.48%. Figure 2 shows the detailed result of bagging, boosting,
dagging, and DECORATE classifiers according to their accuracy. Here, DS is used as a base classifier.

Figure 2. Performance of meta classifiers and voting classifier for ArgoUML project

Int J Artif Intell, Vol. 14, No. 1, February 2025: 376–384

Int J Artif Intell ISSN: 2252-8938 ❒ 381

Performance of voting classifier: in this work, two combination rules of voting are used; average
voting and majority voting. The sub-models of voting classifiers are decided on the individual performance of
the classifiers. For example, in the ArgoUML project, NB, DT, logistic, KLR, NBTree, and MLPCL classifier’s
performance is worse than other classifiers. So, they are used as the base classifier of voting for the ArgoUML
project. In this work, some combinations of base classifiers are used to evaluate the result. First, the result
of NB and DT is noted down. Then classifiers are added to check the performance of the Voting classifier
one by one. Figure 2 reports the detailed result of the voting classifier according to their base classifiers. The
accuracy of NB and DT are 79.37% and 90.3% respectively. When it is used as the sub-models of the voting
classifier, the accuracy has been increased to 91.68% (both for average and majority voting). In the Xerces
project, NB performs poorly than other classifiers in terms of accuracy. Other classifiers like DT, logistic,
KLR, and bagging cannot provide satisfactory results either. So, the different combinations of classifiers are
taken as a base classifier for the voting meta classifier to evaluate the result. When NB and DT are taken as
a combination of base classifiers, the accuracy has been increased i.e., 90.61%. From this analysis, it can be
said that voting classifiers can increase the performance (proper combination of base classifiers) than individual
classifiers. Table 2 shows the result of the classifiers in terms of recall and accuracy.

Table 2. Recall and accuracy of classifiers with and without DPP for ArgoUML project
Classifiers Recall (without DPP) Recall (with DPP) Accuracy (without DPP) Accuracy (with DPP)
SVM 0.925 0.948 0.921 0.925
NB 0.739 0.833 0.739 0.793
RF 0.926 0.949 0.915 0.927
DT 0.903 0.935 0.901 0.903
LibSVM 0.954 0.964 0.953 0.965
Logistic 0.899 0.932 0.899 0.907
KLR 0.918 0.943 0.912 0.918
NBTree 0.931 0.941 0.875 0.88
MLP CL 0.916 0.943 0.916 0.918
RAB 0.919 0.945 0.918 0.923
Bagging 0.916 0.943 0.915 0.917
Dagging 0.981 0.983 0.974 0.983
LB 0.918 0.944 0.908 0.918
DECORATE 0.925 0.949 0.925 0.929
DS 0.913 0.942 0.913 0.913

In comparison with the DPP, we found most of the classifiers perform well. Dagging classifier per-
forms better than other classifiers in terms of accuracy i.e., 97.4% whereas the accuracy of dagging is 98.3%
with the DPP. Maiga et al. [5], [6] proposed two SVM-based works SMURF and SVMDetect. The recall
of SMURF and SVMDetect is 84.09%. Whereas after using DPP the recall of SVM enhances to 94.8% in
this work. Furthermore, from Table 3 we can see that, DPP is increasing the recall of the classifiers to detect
anti-patterns.

Table 3. Comparison of proposed word with existing models in terms of recall

Authors Model Recall (%)
Maiga et al. [5] SMURF 84.09
Maiga et al. [6] SVMDetect 84.09
Proposed method SVM with DPP 94.8

4.1.2. RQ2: How do the classifiers perform in terms of time?

In this work, sixteen different classifiers are taken into consideration. Most of the classifiers performed
well in terms of accuracy and time. NB, DT, SVM, LibSVM, logistic, and LB have taken less time to build a
model. They have taken 0.01 seconds to create the model. NBTree has taken the most time to build the model
which is 0.68 seconds. The voting classifier also takes more time when NBTree is added with other classifiers.
From the performance analysis of different classifiers on different projects, it can be said that NBTree performs
poorly in terms of time than other classifiers. Figure 3 shows the detailed result of the time that is taken by the
classifiers to build the model.

A systematic analysis on machine learning classifiers with data pre-processing to ... (Nazneen Akhter)
382 ❒ ISSN: 2252-8938

Figure 3. Time complexity to build the model for ArgoUML project

5. CONCLUSION
Anti-pattern is a significant nuisance in source code which violates the standard of design pattern.
Software quality makes the software understandable, reusable, and cost-effective. Hence, anti-pattern detection
has become a major concern. However, there are many difficulties in choosing appropriate classifiers to identify
the design pattern. In this paper, a detailed discussion about 16 different well-known classifiers is done to
detect anti-patterns. This analysis will help practitioners to choose suitable classifiers. In addition, to improve
the performance of the classifiers, SMOTE DPP is used. Our classifiers performed better with DPP than the
existing models without DPP. Dagging, KLR, and LibSVM perform better in terms of accuracy for respective
projects. The accuracy of them is 98.4%, 97%, and 95.83% respectively. However, this work presents better
accuracy than existing models. However, the data size used for this work is small. In the future, the dataset
size can be increased by adding more samples. The data set can be applied to a deep learning model adopting
an attention mechanism.

ACKNOWLEDGEMENTS
This work was supported by the Woosong University Academic Research Fund, 2024, South Korea.

REFERENCES
[1] G. Rasool and Z. Arshad, “A review of code smell mining techniques,” Journal of Software: Evolution and Process, vol. 27, no. 11,
pp. 867–895, 2015, doi: 10.1002/smr.1737.
[2] F. Palomba, G. Bavota, M. D. Penta, R. Oliveto, A. D. Lucia, and D. Poshyvanyk, “Detecting bad smells in source code using
change history information,” in 2013 28th IEEE/ACM International Conference on Automated Software Engineering (ASE), 2013,
pp. 268–278, doi: 10.1109/ASE.2013.6693086.
[3] M. Fowler, K. Beck, J. Brant, W. Opdyke, D. Roberts, and E. Gamma, Refactoring: improving the design of existing code. New
York, United States of America: Pearson Education, 1999.
[4] N. Akhter, S. Rahman, and K. A. Taher, “An anti-pattern detection technique using machine learning to improve code quality,” in
2021 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD), 2021,
pp. 356–360, doi: 10.1109/ICICT4SD50815.2021.9396937.
[5] A. Maiga, N. Ali, N. Bhattacharya, A. Sabane, Y. -G. Gueheneuc, and E. Aimeur, “SMURF: A SVM-based incremental anti-pattern
detection approach,” in 2012 19th Working Conference on Reverse Engineering, 2012, pp. 466–475, doi: 10.1109/WCRE.2012.56.
[6] A. Maiga et al., “Support vector machines for anti-pattern detection,” in 2012 27th IEEE/ACM International Conference on Auto-
mated Software Engineering, 2012, pp. 278–281, doi: 10.1145/2351676.2351723.
[7] F. Pecorelli, F. Palomba, D. D. Nucci, and A. D. Lucia, “Comparing heuristic and machine learning approaches for metric-based code
smell detection,” in IEEE International Conference on Program Comprehension, 2019, pp. 93–104, doi: 10.1109/ICPC.2019.00023.
[8] F. A. Fontana, M. V. Mäntylä, M. Zanoni, and A. Marino, “Comparing and experimenting machine learning techniques for code
smell detection,” Empirical Software Engineering, vol. 21, no. 3, pp. 1143–1191, 2016, doi: 10.1007/s10664-015-9378-4.
[9] M. Kessentini, H. Sahraoui, M. Boukadoum, and M. Wimmer, “Search-based design defects detection by example,” in Fundamental
Approaches to Software Engineering, Berlin, Heidelberg, 2011, pp. 401–415, doi: 10.1007/978-3-642-19811-3 28.
[10] I. Polášek, S. Snopko, and I. Kapustik, “Automatic identification of the anti-patterns using the rule-based approach,” in 2012 IEEE
10th Jubilee International Symposium on Intelligent Systems and Informatics, 2012, pp. 283–286, doi: 10.1109/SISY.2012.6339530.
[11] M. T. Aras and Y. E. Selcuk, “Metric and rule based automated detection of antipatterns in object-oriented software sys-
tems,” in 2016 7th International Conference on Computer Science and Information Technology (CSIT), 2016, pp. 1–6, doi:
10.1109/CSIT.2016.7549470.
[12] S. Uchiyama, H. Washizaki, Y. Fukazawa, and A. Kubo, “Design pattern detection using software metrics and machine learning,”
First international workshop on model-driven software migration (MDSM 2011), vol. 708, pp. 38–47, 2011.
[13] A. Barbez, F. Khomh, and Y.-G. Guéhéneuc, “A machine-learning based ensemble method for anti-patterns detection,” Journal of
Systems and Software, vol. 161, 2020, doi: 10.1016/j.jss.2019.110486.
[14] Y. Yin, Q. Su, and L. Liu, “Software smell detection based on machine learning and its empirical study,” in Second Target Recogni-
tion and Artificial Intelligence Summit Forum, 2020, doi: 10.1117/12.2550500.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 376–384

Int J Artif Intell ISSN: 2252-8938 ❒ 383

[15] U. Azadi, F. A. Fontana, and M. Zanoni, “Machine learning based code smell detection through WekaNose,” in Proceed-
ings of the 40th International Conference on Software Engineering: Companion Proceeedings, 2018, pp. 288–289, doi:
10.1145/3183440.3194974.
[16] D. Cruz, A. Santana, and E. Figueiredo, “Detecting bad smells with machine learning algorithms,” in Proceedings of the 3rd
International Conference on Technical Debt, 2020, pp. 31–40, doi: 10.1145/3387906.3388618.
[17] L. Kumar and A. Sureka, “An empirical analysis on web service anti-pattern detection using a machine learning framework,” in
2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC), 2018, pp. 2–11, doi: 10.1109/COMP-
SAC.2018.00010.
[18] N. Pritam et al., “Assessment of code smell for predicting class change proneness using machine learning,” IEEE Access, vol. 7, pp.
37414–37425, 2019, doi: 10.1109/ACCESS.2019.2905133.
[19] A. Jesudoss, S. Maneesha, and T. L. N. Durga, “Identification of code smell using machine learning,” in 2019 International Confer-
ence on Intelligent Computing and Control Systems (ICCS), 2019, pp. 54–58, doi: 10.1109/ICCS45141.2019.9065317.
[20] M. H. -Kacem and N. Bouassida, “A hybrid approach to detect code smells using deep learning,” in Proceedings of
the 13th International Conference on Evaluation of Novel Approaches to Software Engineering, 2018, pp. 137–146, doi:
10.5220/0006709801370146.
[21] A. K. Das, S. Yadav, and S. Dhal, “Detecting code smells using deep learning,” in IEEE Region 10 Annual International Conference,
Proceedings/TENCON, 2019, pp. 2081–2086, doi: 10.1109/TENCON.2019.8929628.
[22] A. Barbez, F. Khomh, and Y.-G. Gueheneuc, “Deep learning anti-patterns from code metrics history,” in 2019 IEEE International
Conference on Software Maintenance and Evolution (ICSME), 2019, pp. 114–124, doi: 10.1109/ICSME.2019.00021.
[23] R. K. Dash, “Selection of the best classifier from different datasets using WEKA,” International Journal of Engineering Research
& Technology (IJERT), vol. 2, no. 3, pp. 1–7, 2013.
[24] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,” Journal
of Artificial Intelligence Research, vol. 16, pp. 321–357, 2002, doi: 10.1613/jair.953.
[25] V. Jakkula, “Tutorial on support vector machine (SVM),” School of EECS, Washington State University, vol. 37, pp. 1–13, 2011.
[26] W. Chen, X. Yan, Z. Zhao, H. Hong, D. T. Bui, and B. Pradhan, “Spatial prediction of landslide susceptibility using data mining-
based kernel logistic regression, naive Bayes and RBFNetwork models for the Long County area (China),” Bulletin of Engineering
Geology and the Environment, vol. 78, no. 1, pp. 247–266, 2019, doi: 10.1007/s10064-018-1256-z.
[27] C. C. Chang and C. J. Lin, “LIBSVM: A library for support vector machines,” ACM Transactions on Intelligent Systems and
Technology, vol. 2, no. 3, 2011, doi: 10.1145/1961189.1961199.
[28] S. L. Cessie and J. C. V. Houwelingen, “Ridge estimators in logistic regression,” Applied statistics, vol. 41, no. 1, pp. 191-201, 1992.
[29] I. Rish, “An empirical study of the naive bayes classifier,” IJCAI 2001 workshop on empirical methods in artificial intelligence, vol.
3, no. 22, pp. 41–46, 2001.
[30] R. Kohavi, “Scaling up the accuracy of naive-Bayes classifiers: A decision-tree hybrid,” in 2nd International Conference on Knowl-
edge Discovery and Data Mining, KDD 1996, 1996, pp. 202–207.
[31] T. M. Christian and M. Ayub, “Exploration of classification using NBTree for predicting students’ performance,” in 2014 Interna-
tional Conference on Data and Software Engineering (ICODSE), 2014, pp. 1–6, doi: 10.1109/ICODSE.2014.7062654.
[32] J. R. Quinlan, C4.5: Program for machine learning. San Mateo, California: Morgan Kaufmann Publishers, 1993.
[33] C. Sammut and G. I. Webb, “Decision stump,” in Encyclopedia of Machine Learning, Boston, MA: Springer, 2011, pp. 262–263,
doi: 10.1007/978-0-387-30164-8 202.
[34] L. Breiman, “Random forests,” in Machine Learning, Springer, 2001, pp. 5–32, doi: 10.1023/A:1010933404324.
[35] B. Ziółko, S. Manandhar, R. C. Wilson, and M. Ziółko, “Logitboost WEKA classifier speech segmentation,” in 2008 IEEE Interna-
tional Conference on Multimedia and Expo, 2008, pp. 1297–1300, doi: 10.1109/ICME.2008.4607680.
[36] J. Friedman, T. Hastie, and R. Tibshirani, “Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder
by the authors),” The Annals of Statistics, vol. 28, no. 2, pp. 337–407, 2000, doi: 10.1214/aos/1016218223.
[37] L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp. 123–140, 1996, doi: 10.1023/A:1018054314350.
[38] P. Roy, R. Chakrabortty, I. Chowdhuri, S. Malik, B. Das, and S. C. Pal, “Development of different machine learning ensemble
classifier for gully erosion susceptibility in Gandheswari Watershed of West Bengal, India,” in Machine Learning for Intelligent
Decision Science, 2020, pp. 1–26, doi: 10.1007/978-981-15-3689-2 1.
[39] J. Kittler, M. Hatef, R. P. W. Duin, and J. Matas, “On combining classifiers,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 20, no. 3, pp. 226–239, 1998, doi: 10.1109/34.667881.
[40] S. Srivastava, “Weka: a tool for data preprocessing, classification, ensemble, clustering, and association rule mining,” International
Journal of Computer Applications, vol. 88, no. 10, pp. 26–29, 2014, doi: 10.5120/15389-3809.

BIOGRAPHIES OF AUTHORS

Nazneen Akhter completed her B.Sc. in Information and Communication Engineering

and M.Sc. in Information and Communication Engineering from the Department of Information and
Communication Technology of Bangladesh University of Professionals, besides several professional
skills. She joined the Department of Information and Communication Technology as a lecturer in
2021. She is currently serving as a lecturer in the Department of Computer Science and Engineering
at Bangladesh University of Professionals, Dhaka, Bangladesh. Her research areas of interest include
software engineering, artificial intelligence and applications, and natural language processing. She
can be contacted at email: nazneen.akhter@bup.edu.bd.

A systematic analysis on machine learning classifiers with data pre-processing to ... (Nazneen Akhter)
384 ❒ ISSN: 2252-8938

Afrina Khatun obtained her B.Sc. in Software Engineering and M.Sc. in Software
Engineering from the Institute of Information Technology, University of Dhaka. She joined the
Department of Information and Communication Technology as a lecturer in 2021. Currently, she
is working as an Assistant Professor at the Bangladesh University of Professionals in the Department
of Computer Science and Engineering. Her research interest is the application of machine learning
and deep learning in software engineering. She can be contacted at email: afrina.khatun@bup.edu.bd.

Dr. Md. Sazzadur Rahman is working as a Professor at the Institute of Information

Technology (IIT), Jahangirnagar University, Savar, Dhaka-1342, Bangladesh. He received his Ph.D.
in Material Science (Surface Science) from Kyushu University, Japan as a MEXT scholarship grantee
in 2015. Previously, he received his B.Sc. (Hons.) degree and M.Sc. degree in Applied Physics, Elec-
tronics, and Communication Engineering from the University of Dhaka, Bangladesh. He has about
14 years of experience in research and teaching in electronics, internet of things, machine learning,
communication engineering, and material science. Now he is working as a co-investigator on the DI-
VERSASIA Erasmus+ project (project number 618615-EPP-1-2020-1-UKEPPKA2-CBHEJP). He
can be contacted at email: sazzad@juniv.edu.

A. S. M. Sanwar Hosen received the M.S. and Ph.D. degrees in computer science and
engineering from Jeonbuk National University, Jeonju, South Korea. He is currently an Assistant
Professor at the Department of Artificial Intelligence and Big Data, Woosong University, Daejeon,
South Korea. He has published several articles in journals and international conferences. His current
research interests include wireless sensor networks, the internet of things, fog-cloud computing, cyber
security, artificial intelligence, and blockchain technology. He has been an expert reviewer for IEEE
Transactions, Elsevier, Springer, and MDPI journals and magazines. He has also been invited to serve
as the Technical Program Committee Member for several reputed international conferences, such as
IEEE ACM. He can be contacted at email: sanwar@wsu.ac.kr or sanwar@jbnu.ac.kr.

Mohammad Shahidul Islam received his Ph.D. in Computer Science and Information
Systems from the National Institute of Development Administration (NIDA), Bangkok, Thailand,
B.Tech. in Computer Science and Technology from the Indian Institute of Technology-Roorkee
(IITR), Uttar Pradesh, India in 2002, M.Sc. in Mobile Computing and Communication from Uni-
versity of Greenwich, London, U.K in 2008. His research interests include artificial intelligence,
robotics, machine learning, pattern recognition, and renewable energy. He can be contacted at email:
suva.93@grads.nida.ac.th or suva93@gmail.com.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 376–384