In recent decades, electronic communication became one
of the most important communication channels. Informal in-
person chats were replaced by online forums, posting on
social media and direct messaging. This opened the doors
of a new kind of harassment, online harassment [1]. Unlike
few decades ago, when harassment was contained in either
direct face-to-face insults or written communication (letters,
articles, street graffiti,etc.), now it is mainly conducted on
online platforms. This is why now it can have much broader
reach, and combined with the anonymity of online landscape,
but also it can keep individuals anonymous and therefore
not accountable. This can have a disastrous effect on more
vulnerable people and groups, especially younger ones who
did not develop appropriate coping mechanisms [2].
It is important to conduct research and development in the
field of detection and prevention mechanisms that will keep
sure our online world more friendly, decent and less toxic.
In general, detection is difficult due to two main reasons.
First, online platforms generate vast data streams each second,
so these mechanisms need to process large amounts of data.
Second difficulty is the constant change of online language.
All of these online trends (e.g. TikTok or Instagram trends and
slangs) change rapidly [3], and each one of them opens new
types of potential harassment. Since the culture changes quite
fast, harassment takes new form all the time - which in return
makes designing detection systems.
This is why the recent development of artificial intelligence
(AI) comes very handy. Instead of relying on explicit instruc-
tions, AI systems can adopt themselves with new data. That is
what makes them suitable for dealing with constantly changing
ways of online harassment. In general AI models consist of
architecture and hyperparameters that are chosen while they
are designed and parameters that are tuned during training
and are constantly updated with the flow of new data. In this
paper the problem is tackled with natural language processing
(NLP) [4], where the selection of the hyperparameters is an
NP-hard problem. This is why metaheuristic is used concretely
Modified Metaheuristics Optimization for Cyberbullying Detection on
Online Data Science Platform
NEBOJSA BACANIN1a, LUKA JOVANOVIC2b, ILJA UZELAC BUJISIC2,
JELENA KALJEVIC3c, JELENA CADJENOVIC4d, MILOS ANTONIJEVIC1e,
MIODRAG ZIVKOVIC1f
1Faculty of Informatics and Computing
Singidunum University
Belgrade, SERBIA
2Faculty of Technical Sciences
Singidunum University
Belgrade, SERBIA
3Faculty for Health and Business Studies
Singidunum University
Belgrade, SERBIA
4Deprtmant of Business Economics
Singidunum University
Belgrade, SERBIA
a0000-0002-2062-924X, b0000-0001-9402-7391, c0000-0002-5135-8083,
d0000-0002-3406-3347, e0000-0002-5511-2531, f0000-0002-4351-068X
Abstract: —Online harassment detection faces significant challenges due to its expansive reach and anonymity. Addressing
this issue demands effective detection mechanisms capable of processing vast data streams and adapting to evolving online
language. Leveraging advancements in artificial intelligence, we propose a novel approach grounded in natural language
processing (NLP) and metaheuristic algorithms. Our methodology integrates term frequency-inverse document frequency
(TF-IDF) encoding and the AdaBoost algorithm for classification. To tackle the NP-hard problem of hyperparameter
selection, we introduce a modified crayfish optimization algorithm (COA), termed GI-COA. This paper represents a
pioneering effort in utilizing metaheuristic algorithms for hyperparameter selection in harassment detection models.
Through experimentation, we demonstrate the efficacy of our approach in fostering a safer online environment. The best
performing optimize models demonstrate and accuracy exceeding 77%.
Key-words:—Online harassment, Hyperparameter optimization, Metaheuristic algorithms, Natural language processing,
Hybridization,
Received: March 2, 2024. Revised: September 6, 2024. Accepted: October 4, 2024. Published: November 5, 2024.
1. Introduction
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
211
Volume 23, 2024
crayfish optimisation algorithm (COA) [5] in the hyperparam-
eter selection process.
This paper tackles this problem using few novel approaches.
The first question is how to handle data, i.e. how to code
words and their appearances in documents. To grasp this,
term frequency - inverse document frequency (TF-IDF) [6] is
implemented as the main metric. The classifier that classifies
documents as harassment or non-harassment is based on Ad-
aBoost [7] algorithm. The NP-hard problem of hyperparameter
tuning is tackled by a modification of COA. This paper is a first
bridge of using metaheuristical algorithms in hyperparameter
selection in models for harassment detection.
The rest of the paper starts with the overview of the general
works in NLP models and existing spam detection models.
This is followed by a short discussion about metaheuristic
algorithms in general. Brief descriptions of AdaBoost, TF-IDF,
COA and the modified version of COA, GI-COA are given.
After the experimental setup is covered, simulation outcomes
are obtained and discussed before the conclusion of the paper.
In general, work in harassment detection lies closely to spam
detection problems as a part of wider NLP research. In this
broad and fast-developing area some of the most important
research papers are [8] introducing attention-based transformer
models, [9] which introduced the original GPT model and
demonstrated the effectiveness of generative pre-training on a
diverse set of NLP tasks and [10] which introduced the more
powerful GPT-2 which demonstrated effectiveness on various
NLP tasks to the point that it sparked ethical considerations of
such models.Some relevant papers in spam detection include
BERT for Spam detection [11] which use BERT, bidirectional
encoder representation from transformers introduced in [12].
which is based on support vector machines. The SVM based
approach can be useful due to its effectiveness with dealing
with high-dimensional data but it can be computationally
expensive and requires careful selection of hyperparameters
and kernel functions which might be impractical. On the
other hand BERT based approaches show strong results on
vast NLP tasks including spam, and large text corpus it was
trained on makes it very generalizable. On the other hand, its
large number of hyperparameters makes the training process
expensive.
When using machine learning algorithms, some architecture
is chosen and then parameters are trained using learning data.
The difficulty of the process lies in selecting an appropriate
model, finding and processing training data and then tuning the
parameters using it. However, before the training process, it is
often necessary to define some parameters that will not change
during the training but will significantly affect the performance
of the model. This values are referred to as hyperparameters.
The process of finding appropriate hyperparameters consists of
determining which parameters are hyperparameters, defining
the search range for each of them, choosing a metric to
grade each choice, defining the search strategy, tuning the
hyperparameters, selecting the best choice and then testing
the model with these hyperparameters on some independent
test set. Since grid search if done directly can be an NP-hard
problem, a metaheuristical approach is used to find a good
choice of hyperparameters.
Literature presents many potential candidate algorithms
of optimizations. Some popular choices used by researchers
include established optimizers like the variable neighborhood
search (VNS) [13], GA [14], and sine cosine algorithms
(SCA) [15] are also examined. Metaphor-based optimizers,
including bat algorithm (BA) [16], whale optimization algo-
rithm (WOA) [17], and Harris hawk optimizer (HHO) [18],
are also popular choices. Some newer examples include the
Botox optimization algorithm BOA [19] inspired by medical
treatment processes.
Optimizes have been successfully applied in several field
such as cybersecurity where these algorithms tackle intrusion
detection [20], spam identification [21] as well as fraud iden-
tification [22]. However, hybrid algorithms have also proven
to be highly effective at tackling optimizations with notable
implementations showing interesting results in finance [23]
and medical fields [24]. With more recent examples tackling
problems in energy [25] and production forecasting [26].
The AdaBoost [7] optimizer widely considered a good
choice for optimization problems. This approach adopts en-
semble techniques and uses several simpler models to for-
mulate a decision. By applying weights to classifier votes
better decisions can be made though a sort of consensus
method. Should a correct classification be made weights of
classifiers are increased, otherwise weights are decreased.
Classifier errors can be determined as per Eq.(1):
ϵt=PN
i=1 wi,t ·I(ht(xi)=yi)
PN
i=1 wi,t
(1)
the term ϵtrepresents a weighted adjustment for a weak model,
where wi,t corresponds to the weight. The expression ht(xi)
refers to the prediction made by the weak classifier, and yi
indicates the actual value or ground truth. The function I
converts the output to 1 if the classification is positive, and
0 if it is negative.
Additional classifiers are built using the determined weights,
and this weight-adjustment process is repeated. Generally, a
large number of classifiers are accumulated to create accurate
models. Each sub-model is given a score, and these are
combined to form a linear model. The classifier’s weight
within the ensemble can be computed using Eq. 2.
αt=1
2ln 1−ϵt
ϵt(2)
the term αtsignifies the assigned weight for each weak
classifier. This value is determined based on the error value,
ϵt, and defines the classifier’s contribution to the final result.
These weights are updated according to the following rule:
wi,t+1 =wi,t ·exp (−αt·yi·ht(xi)) (3)
2. Related Works
2.1 Adaboost
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
212
Volume 23, 2024
where wi,t represents the weight a given instance, αtdenotes
the weight of the weak classifier yiis the ground truth value,
and ht(xi)signifies the prediction.
It is difficult to directly convert words into numbers in
a meaningful way. What matters is to somehow quantify
sentences such that from that quantification there is some
insight about it. To do that, a measure called TF-IDF is
applied. That measure captures how relevant is each word in
a document where a document belongs to some collection of
documents - i.e. a corpus.
Intuitively speaking useful measure would be a one that
increases proportionally with each new appearance of a word
in a document, and also decreases with the frequency of
word appearance in corpus. This makes sense, since to extract
relevant words for a specific documents, the words that appear
in many documents should be penalised. These words will
not give any insight of this specific document that is being
analysed. Formally for a word w, document din a corpus C
of documents TF and IDF are:
TF =num of appearances of win d
num of words in d(4)
IDF = log2num of documents in C
num of documents in Ccontaining w(5)
Then TF-IDF is defined by TF-IDF := TF*IDF. This is the
quantity that will be used for relevant words. Note that the
base of logarithm is not important since it just scales TF-IDF
by a constant. More on TF-IDF can be found in [27]
Crayfish, also known as crawfish or freshwater lobsters, are
crustecans resembling small lobsters that typically live in river
waters. The COA is a metaheuristic algorithm inspired by
behaviour of crayfish. Behaviour consists of foraging, com-
petition and summer vacation stage. First two stages are the
exploitation stage while summer vacation is exploration stage
of COA. First all the variables used in COA are explained. Let
nbe the number of crayfishes, Tthe number of iterations ,d
the dimension of space in which the process takes place. The
position of crayfishes at time t, for 0≤t≤Tis an n×dmatrix
X(t)where X(t)
iis the d-dimensional vector representing the
location of i-th crayfish at time tfor 1≤i≤n.
The n×dmatrices Lband Ubare lower and upper bounds
on position of crayfishes. At each time, the temperature temp
of the system is updated. Temperature defines in which of
the three possible states (foraging, competition and summer
vacation) the system is which is essential for knowing the
next position. The location of food (in cases when used)
will be Xfood,Xshade the location of the cave, fthe fitness
function, that grades each position and pis a random variable
depending on temp. More details on these variables can be
found in [28]. The initial positions of crayfishes is chosen by
X(0) =Lb+(Ub−Lb)∗rand where rand is a random variable
uniformly distributed in [0,1]. (Essentially a random point is
picked in the d-dimensional space between the given lower
and upper bounds for each crayfish.) In each new time-step,
first the temperature temp is updated. Take:
temp = 20 + 15 ∗rand (6)
where rand is a uniform random variable in [0,1] that is
updated in each step.
Depending on whether temp > 30 or not there are two sub-
cases. Algorithm enters in first case if temp > 30 (summer
resort or competition stage). Then the two cases are separated
: if rand < 0.5it is in summer resort stage, otherwise it is in
competition stage. In this case, crayfish will approach to cave
for summer resort, instead of fighting. Then the values of X
will be updated according to the following equation:
X(t+1)
i,j =X(t)
i,j + (2 −t
T)∗rand ∗(Xshade −X(t)
i,j )(7)
To see how we update Xshade in each iteration, check [28].
This means that crayfishes start fighting for a cave. To capture
the fight, another variable zis introduced:
z:= round(rand ∗(n−1)) + 1 (8)
Then:
X(t+1)
i,j =X(t)
i,j −X(t)
z,j +Xshade.(9)
Algorithm enters second case if temp ≤30. Here, the
crayfish go to food which is located on XG=Xfood. The
food size Qis defined as:
Q= 3 ∗rand ∗f(X(t)
i)
f(Xfood).(10)
Depending on the size of the food, crayfish either splits it
into two parts, or it moves to it and eats it directly. If Q > 2,
then Xfood becomes e
−1
Q∗Xfood and
X(t+1)
i,j =X(t)
i,j +Xfood ∗p∗(cos(2π∗rand)−sin(2π∗rand))
(11)
Otherwise, if Q≤2, then:
X(t+1)
i,j = ((X(t)
i,j −Xfood) + rand ∗X(t)
i,j )∗p(12)
This gives the idea on how Xevolves in time. The final X
at time Tis the solution that is taken as the result of COA for
hyperparameter selection.
This was a brief overview of COA. To see more details, check
[28].
Although the initial COA performs well, it has some short-
comings when evaluated using standard CEC [29] functions.
To address these limitations and enhance COA’s effectiveness,
this study introduces hybridization techniques. Inspired by the
Genetic Algorithm (GA) [14], we propose a new algorithm
called the genetically inspired COA (GICOA).
In the GICOA algorithm, a unique mechanism is activated
after each iteration, selecting a random agent and merging
it with the best solution found so far. Their parameters
2.2 TF-IDF
3.1 The Original Crayfish Optimisation Algorithm
3. Methods
3.2 Generically Inspired COA
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
213
Volume 23, 2024
are uniformly blended using a control parameter denoted as
pc, which is empirically set to 0.1. Additionally, parameter
mutation is introduced: this process selects a random value
within a specified range, then either adds or subtracts half of
this value from the parameter based on the mutation direction
parameter md, also set to 0.1.
After generating a new solution, the worst-performing so-
lution in the swarm is replaced by the new agent. The evalu-
ation of the new solution is deferred until the next iteration,
keeping the computational complexity the same as the original
algorithm. To provide a comprehensive understanding, the
pseudocode is presented in Algorithm 1.
Algorithm 1 Pseudo code of the introduced GICOA.
Generate population P
while current;itteration is less than maxiumum;itterations do
Assess agents in P
for Each agent Xin Pdo
Update locations by applying the COA search
Generate a new solution NS, with based on introduced mechanism
Mutate genome of NS
Replace the worst solution in Pwith NS
end for
end while
return The best solution attained within P
To evaluate the performance of the introduced optimizer and
the overall viability of the introduced approach a modified
real world curated dataset is used acquired form Kaggle 1.
The parse Kaggle parsed portion of the dataset is used for
simulations conducted in this work. Several contemporary
optimizers are subjected to a comparative analysis. The op-
timizers included in the comparisons include the modified
algorithm alongside the original COA [5]. Well established
optimizers are also evaluated such as the VNS [13], GA [14]
and SCA [15]. Further metaphor based optimizers such as the
BA [16], WOA [17] and HHO [18] are evaluated. Finally
the recently introduced BOA [19] are assessed. Algorithms
are independently implemented according to the parameter
settings provided in the original works that introduced the
optimizers.
AdaBoost parameters are subjected to optimization with
constraints provided in Table I. The respective constrains have
been empirically determined. Each optimizer was allocated
ten iterations and a population size of eight agents each.
Optimizations are carried out in 30 independent runs. Standard
classifications metrics are used to evaluate the tested optimizes
including accuracy, precision, f1-score and recall. Finally the
Cohen’s kappa [30] metric is used as the indicator function
due to the imbalance observed in the dataset.
TABLE I
ADABOOST HYPERPARAMETERS SUBJECTED TO OPTIMIZATION.
Parameter constraints
count of estimators [10,20]
depth [1,5]
learning rate [0.01,2]
1https://www.kaggle.com/datasets/saurabhshahane/cyberbullying-dataset
TABLE II
OBJECTIVE FUNCTION OUTCOME FOR THE CONDUCTED SIMULATIONS.
Method Best Worst Mean Median Std Var
AB-GICOA 0.479523 0.464299 0.471482 0.471334 0.005674 3.22E-05
AB-COA 0.463551 0.440244 0.454147 0.455728 0.006455 4.17E-05
AB-VNS 0.464634 0.448140 0.458246 0.458168 0.004878 2.38E-05
AB-GA 0.473523 0.452341 0.460195 0.458979 0.006538 4.27E-05
AB-BA 0.475901 0.444988 0.457535 0.454765 0.009970 9.94E-05
AB-SCA 0.473954 0.451173 0.461481 0.461740 0.007691 5.92E-05
AB-WOA 0.471007 0.450067 0.459053 0.458005 0.005476 3.00E-05
AB-HHO 0.464634 0.453648 0.458976 0.458444 0.003423 1.17E-05
AB-BOA 0.473954 0.453393 0.463898 0.464007 0.006615 4.38E-05
TABLE III
CAPTION
Method Best Worst Mean Median Std Var
AB-GICOA 0.221340 0.226190 0.223821 0.223986 0.001777 3.16E-06
AB-COA 0.227072 0.232363 0.229773 0.229497 0.002061 4.25E-06
AB-VNS 0.226190 0.231922 0.229112 0.229056 0.002434 5.93E-06
AB-GA 0.223545 0.234127 0.229112 0.228395 0.003575 1.28E-05
AB-BA 0.223986 0.233686 0.229497 0.229277 0.003837 1.47E-05
AB-SCA 0.223986 0.233686 0.228560 0.228836 0.003436 1.18E-05
AB-WOA 0.223986 0.236332 0.228671 0.227734 0.003386 1.15E-05
AB-HHO 0.226190 0.235009 0.229497 0.229056 0.002736 7.48E-06
AB-BOA 0.223986 0.233245 0.226852 0.225970 0.002866 8.21E-06
Objective function outcomes for classifiers optimized by
each model are provided in Table II. Outcomes are show-
cased in terms of best, worst, mean and median outcomes.
Furthermore details on algorithms stability are provided in
the form of the standard deviation and variance in outcomes
over 30 independent runs. As shown the introduced optimizer
outperformed competing models across al metrics including
stability.
Indicator function outcomes for classifiers optimized by
each model are provided in Table III. Outcomes are show-
cased in terms of best, worst, mean and median outcomes.
Furthermore details on algorithms stability are provided in
the form of the standard deviation and variance in outcomes
over 30 independent runs. As shown the introduced optimizer
outperformed competing models across all metrics including
stability.
Comparisons in terms of stability are provided in Figure 1.
The introduced modified algorithm outperforms competing
optimizers including the original version, providing a high rate
of stability as well as the highest quality of results.
Further detailed comparisons in terms of accuracy, per-class
precision as well as macro and weight averages are provided
in Table IV. The introduced optimize show the highest rate of
accuracy, as well as macro and weighted averages. However,
certain algorithms outperform the introduced optimizer in
terms of precision and recall. This is to be somewhat expected,
as according to the NFL no single optimizer performs equally
well for all tasks and across all metrics. Hence experimentation
is mandated in order to determine the optimal combination op
problem and optimizer.
Further details on an optimizes ability to avoid local optima,
and sufficiently explore the search space for good solutions
can be discerned form the respective optimizers convergence
graphs. Graphs for tested algorithms are provide in Figure 2.
The introduced optimizer manages to avoid local optima that
constrains other optimizers with insufficient exploration ability
attaining the best comparative outcomes in the ninth iteration.
4. Experimental Setup
5. Experimental Outcomes
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
214
Volume 23, 2024
AB-GICOA
AB-COA
AB-VNS
AB-GA
AB-BA
AB-SCA
AB-WOA
AB-HHO
AB-BOA
Algorithm
0.43
0.44
0.45
0.46
0.47
0.48
0.49
Objective
Cyberbullying Kaggle harassment (TF-IDF) - objective violin plot diagram
AB-GICOA
AB-COA
AB-VNS
AB-GA
AB-BA
AB-SCA
AB-WOA
AB-HHO
AB-BOA
Algorithm
0.222
0.224
0.226
0.228
0.230
0.232
0.234
0.236
error
Cyberbullying Kaggle harassment (TF-IDF) - error box plot diagram
Fig. 1. Objective and indicator function distributions.
TABLE IV
CAPTION
Method metric non-harassment harassment accuracy macro avg weighted avg
AB-GICOA precision 0.785420 0.758377 0.778660 0.771899 0.775941
recall 0.906993 0.540881 0.778660 0.723937 0.778660
f1-score 0.841840 0.631424 0.778660 0.736632 0.768083
AB-COA precision 0.779138 0.753623 0.772928 0.766380 0.770194
recall 0.907671 0.523270 0.772928 0.715471 0.772928
f1-score 0.838507 0.617669 0.772928 0.728088 0.761097
AB-VNS precision 0.778746 0.758242 0.773810 0.768494 0.771558
recall 0.910387 0.520755 0.773810 0.715571 0.773810
f1-score 0.839437 0.617450 0.773810 0.728443 0.761624
AB-GA precision 0.783118 0.756228 0.776455 0.769673 0.773692
recall 0.906993 0.534591 0.776455 0.720792 0.776455
f1-score 0.840516 0.626382 0.776455 0.733449 0.765456
AB-BA precision 0.786350 0.746141 0.776014 0.766245 0.772256
recall 0.899525 0.547170 0.776014 0.723347 0.776014
f1-score 0.839139 0.631350 0.776014 0.735244 0.766303
AB-SCA precision 0.784325 0.751313 0.776014 0.767819 0.772754
recall 0.903598 0.539623 0.776014 0.721610 0.776014
f1-score 0.839748 0.628111 0.776014 0.733930 0.765563
AB-WOA precision 0.781341 0.759494 0.776014 0.770417 0.773683
recall 0.909708 0.528302 0.776014 0.719005 0.776014
f1-score 0.840652 0.623145 0.776014 0.731899 0.764410
AB-HHO precision 0.778746 0.758242 0.773810 0.768494 0.771558
recall 0.910387 0.520755 0.773810 0.715571 0.773810
f1-score 0.839437 0.617450 0.773810 0.728443 0.761624
AB-BOA precision 0.784325 0.751313 0.776014 0.767819 0.772754
recall 0.903598 0.539623 0.776014 0.721610 0.776014
f1-score 0.839748 0.628111 0.776014 0.733929 0.765563
support 1473 795
0246810
Iterations
0.445
0.450
0.455
0.460
0.465
0.470
0.475
0.480
Objective
Cyberbullying Kaggle harassment (TF-IDF) - objective convergence graphs
AB-GICOA
AB-COA
AB-VNS
AB-GA
AB-BA
AB-SCA
AB-WOA
AB-HHO
AB-BOA
Fig. 2. Objective and indicator function convergence plots.
TABLE V
OPTIMIZER MODEL PARAMETER SELECTIONS MADE BY OPTIMIZERS.
Methods Number of estimators Depth Learning rate
AB-GICOA 19 2 1.327088
AB-COA 16 2 1.267089
AB-VNS 20 2 1.185109
AB-GA 20 2 1.377726
AB-BA 20 2 1.371097
AB-SCA 20 2 1.320228
AB-WOA 17 2 1.346582
AB-HHO 17 2 1.332248
AB-BOA 20 2 1.324441
Additional details on the performance of the best perform-
ing model optimized by the introduced modified algorithm
mare produced in Figure 3 where the confusion matrix and
PR curved are shown. Parameter selections made by each
optimizer for the respective best performing models are also
provided to encourage experimental repeatability in Table V.
Detecting online harassment poses substantial hurdles due
to the wide-ranging scope and the anonymity it often involves.
Overcoming these challenges requires robust detection meth-
ods capable of handling large volumes of data and adapting
to the ever-evolving nature of online communication. Through
the fusion of NLP and metaheuristic algorithms, particularly
leveraging TF-IDF encoding and AdaBoost algorithm, this
work proposes an data driven approach. To address the com-
plexity of hyperparameter selection, a further contribution is
given in the form of a introduced modification to the COA,
dubbed GICOA. Experimental outcomes conducted with real
world data suggest that the best-performing models achieving
an accuracy surpassing 77%, thus contributing significantly to
fostering a safer online ecosystem.
6. Conclusion
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
215
Volume 23, 2024
0.4 0.5 0.6 0.7 0.8 0.9 1.0
recall
0.0
0.2
0.4
0.6
0.8
1.0
precision
AB-GICOA Cyberbullying Kaggle harassment (TF-IDF) - PR curve
no harassment AP:0.784
harassment AP:0.620
micro AP: 0.749
no harassment
harassment
Predicted label
no harassment
harassment
True label
0.907 0.093
0.459 0.541
AB-GICOA Cyberbullying Kaggle harassment (TF-IDF) - confusion matrix
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Fig. 3. Best performing model PR plot and confusion matrix.
Future works will focus on further refining the proposed
approach. Additionally, further applications for the proposed
optimizer will be explored.
[1] R. Kowalski, “Cyberbullying,” in The Routledge international handbook
of human aggression, pp. 131–142, Routledge, 2018.
[2] M. O. Lwin, B. Li, and R. P. Ang, “Stop bugging me: An examination
of adolescents’ protection behavior against online harassment,” Journal
of adolescence, vol. 35, no. 1, pp. 31–41, 2012.
[3] A. E. Mackey, Evaluating the evolution of the English language as seen
in TikTok slang. PhD thesis, Wichita State University, 2023.
[4] K. Chowdhary and K. Chowdhary, “Natural language processing,”
Fundamentals of artificial intelligence, pp. 603–649, 2020.
[5] H. Jia, H. Rao, C. Wen, and S. Mirjalili, “Crayfish optimization
algorithm,” Artificial Intelligence Review, vol. 56, no. Suppl 2, pp. 1919–
1979, 2023.
[6] H. Christian, M. P. Agus, and D. Suhartono, “Single document automatic
text summarization using term frequency-inverse document frequency
(tf-idf),” ComTech: Computer, Mathematics and Engineering Applica-
tions, vol. 7, no. 4, pp. 285–294, 2016.
[7] T. Hastie, S. Rosset, J. Zhu, and H. Zou, “Multi-class adaboost,”
Statistics and its Interface, vol. 2, no. 3, pp. 349–360, 2009.
[8] A. Waswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez,
L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NIPS, 2017.
[9] A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al., “Improving
language understanding by generative pre-training,” 2018.
[10] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al.,
“Language models are unsupervised multitask learners,” OpenAI blog,
vol. 1, no. 8, p. 9, 2019.
[11] V. S. Tida and S. Hsu, “Universal spam detection using transfer learning
of bert model,” arXiv preprint arXiv:2202.03480, 2022.
[12] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training
of deep bidirectional transformers for language understanding,” arXiv
preprint arXiv:1810.04805, 2018.
[13] N. Mladenovi´
c and P. Hansen, “Variable neighborhood search,” Com-
puters & operations research, vol. 24, no. 11, pp. 1097–1100, 1997.
[14] S. Mirjalili and S. Mirjalili, “Genetic algorithm,” Evolutionary algo-
rithms and neural networks: Theory and applications, pp. 43–55, 2019.
[15] S. Mirjalili, “Sca: a sine cosine algorithm for solving optimization
problems,” Knowledge-based systems, vol. 96, pp. 120–133, 2016.
[16] X.-S. Yang and A. Hossein Gandomi, “Bat algorithm: a novel approach
for global engineering optimization,” Engineering computations, vol. 29,
no. 5, pp. 464–483, 2012.
[17] S. Mirjalili and A. Lewis, “The whale optimization algorithm,” Advances
in engineering software, vol. 95, pp. 51–67, 2016.
[18] A. A. Heidari, S. Mirjalili, H. Faris, I. Aljarah, M. Mafarja, and
H. Chen, “Harris hawks optimization: Algorithm and applications,”
Future generation computer systems, vol. 97, pp. 849–872, 2019.
[19] M. Hub´
alovsk´
a, ˇ
S. Hub´
alovsk`
y, and P. Trojovsk`
y, “Botox optimization
algorithm: A new human-based metaheuristic algorithm for solving
optimization problems,” Biomimetics, vol. 9, no. 3, p. 137, 2024.
[20] M. Zivkovic, L. Jovanovic, M. Ivanovic, N. Bacanin, I. Strumberger, and
P. M. Joseph, “Xgboost hyperparameters tuning by fitness-dependent
optimizer for network intrusion detection,” in Communication and
intelligent systems: Proceedings of ICCIS 2021, pp. 947–962, Springer,
2022.
[21] M. Salb, L. Jovanovic, M. Zivkovic, E. Tuba, A. Elsadai, and N. Ba-
canin, “Training logistic regression model by enhanced moth flame
optimizer for spam email classification,” in Computer networks and
inventive communication technologies: Proceedings of fifth ICCNCT
2022, pp. 753–768, Springer, 2022.
[22] A. Petrovic, M. Antonijevic, I. Strumberger, L. Jovanovic, N. Savanovic,
and S. Janicijevic, “The xgboost approach tuned by tlb metaheuristics
for fraud detection,” in Proceedings of the 1st international conference
on innovation in information technology and business (ICIITB 2022),
vol. 104, p. 219, Springer Nature, 2023.
[23] N. Bacanin, M. Zivkovic, L. Jovanovic, M. Ivanovic, and T. A. Rashid,
“Training a multilayer perception for modeling stock price index predic-
tions using modified whale optimization algorithm,” in Computational
Vision and Bio-Inspired Computing: Proceedings of ICCVBIC 2021,
pp. 415–430, Springer, 2022.
[24] L. Jovanovic, M. Zivkovic, M. Antonijevic, D. Jovanovic, M. Ivanovic,
and H. S. Jassim, “An emperor penguin optimizer application for
medical diagnostics,” in 2022 IEEE Zooming Innovation in Consumer
Technologies Conference (ZINC), pp. 191–196, IEEE, 2022.
[25] N. Bacanin, L. Jovanovic, M. Zivkovic, V. Kandasamy, M. Antonijevic,
M. Deveci, and I. Strumberger, “Multivariate energy forecasting via
metaheuristic tuned long-short term memory and gated recurrent unit
neural networks,” Information Sciences, vol. 642, p. 119122, 2023.
[26] L. Jovanovic, D. Jovanovic, N. Bacanin, A. Jovancai Stakic, M. An-
tonijevic, H. Magd, R. Thirumalaisamy, and M. Zivkovic, “Multi-step
crude oil price prediction based on lstm approach tuned by salp swarm
algorithm with disputation operator,” Sustainability, vol. 14, no. 21,
p. 14616, 2022.
[27] H. Christian, M. Agus, and D. Suhartono, “Single document automatic
text summarization using term frequency-inverse document frequency
(tf-idf),” ComTech: Computer, Mathematics and Engineering Applica-
tions, vol. 7, p. 285, 12 2016.
[28] H. Jia, H. Rao, C. Wen, and S. Mirjalili, “Crayfish optimization
algorithm,” Artificial Intelligence Review, vol. 1, p. 1, 09 2023.
[29] J. J. Liang, B. Y. Qu, and P. N. Suganthan, “Problem definitions and
evaluation criteria for the cec 2014 special session and competition
on single objective real-parameter numerical optimization,” Computa-
tional Intelligence Laboratory, Zhengzhou University, Zhengzhou China
and Technical Report, Nanyang Technological University, Singapore,
vol. 635, no. 2, p. 2014, 2013.
[30] M. J. Warrens, “Five ways to look at cohen’s kappa,” Journal of
Psychology & Psychotherapy, vol. 5, 2015.
5HIHUHQFHV
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
216
Volume 23, 2024
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2024.23.20
Nebojsa Bacanin, Luka Jovanovic, Ilja Uzelac Bujisic,
Jelena Kaljevic, Jelena Cadjenovic,
Milos Antonijevic, Miodrag Zivkovic
E-ISSN: 2224-2872
217
Volume 23, 2024