Optimal k-means clustering using artificial bee colony algorithm with variable food sources length

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 5, October 2022, pp. 5435~5443
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i5.pp5435-5443  5435
Journal homepage: http://ijece.iaescore.com
Optimal k-means clustering using artificial bee colony
algorithm with variable food sources length
Sabreen Fawzi Raheem, Maytham Alabbas
Department of Computer Science, College of Computer Science and Information Technology, University of Basrah, Basrah, Iraq
Article Info ABSTRACT
Article history:
Received Aug 17, 2021
Revised May 25, 2022
Accepted Jun 18, 2022
Clustering is a robust machine learning task that involves dividing data
points into a set of groups with similar traits. One of the widely used
methods in this regard is the k-means clustering algorithm due to its
simplicity and effectiveness. However, this algorithm suffers from the
problem of predicting the number and coordinates of the initial clustering
centers. In this paper, a method based on the first artificial bee colony
algorithm with variable-length individuals is proposed to overcome the
limitations of the k-means algorithm. Therefore, the proposed technique will
automatically predict the clusters number (the value of k) and determine the
most suitable coordinates for the initial centers of clustering instead of
manually presetting them. The results were encouraging compared with the
traditional k-means algorithm on three real-life clustering datasets. The
proposed algorithm outperforms the traditional k-means algorithm for all
tested real-life datasets.
Keywords:
Artificial bee colony algorithm
K-means algorithm
Optimize k-means clustering
Variable-length representation
This is an open access article under the CC BY-SA license.
Corresponding Author:
Maytham Alabbas
Department of Computer Science, College of Computer Science and Information Technology,
University of Basrah
Basrah, Iraq
Email: ma@uobasrah.edu.iq
1. INTRODUCTION
It is important to utilize various data mining techniques, including cluster analysis, to identify,
analyze, and categorize data attributes. Researching in data clustering is still active. It is used extensively in
various fields, such as medical sciences, image analysis, machine learning, web cluster engines,
classification, knowledge discovery, and software engineering. Clustering is splitting the area or population
into groups to make data points in the same group more comparable than data points in other groups. It is one
of the most often used strategies for unsupervised classification. Several exist a variety of unsupervised
clustering algorithms available including k-means [1], cobweb [2], farthest-first [3], expectation-
maximization (EM) [4], density-based [5], and hierarchical clustering [6]. On the whole, though, the most
widely used is the k-means algorithm.
K-means clustering is a well-known partition algorithm. It was widely utilized in scientific research
and industrial applications due to its simplicity, rapid convergence, and suitability for massive data sets
processing, among others. The traditional k-means clustering method allocated random beginning points
during clustering center initialization and typically found a local optimum clustering result. As a result, the
lack of stability affected categorization accuracy, and a globally optimized method is required to overcome
the limitations of this algorithm.
Numerous researches have been conducted to address this issue. For instance, Maulik and
Bandyopadhyay [7] suggest using a genetic algorithm (GA) to search in the feature space for cluster centers

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5435-5443
5436
to optimize a similarity measure. The particle swarm optimization (PSO) algorithm was used to locate the
centroids of several clusters that the user specifies [8]–[10]. Kao et al. [11] propose a hybrid method, namely
combining k-means, Nelder-Mead simplex search, and PSO (K–NM–PSO). Niknam and Amiri [12]
suggested the hybrid evolutionary algorithm, fuzzy adaptive particle swarm optimization-ant colony
optimization- k-means algorithms (FAPSO–ACO–K), which has a greater chance of clustering. Laszlo and
Mukherjee [13] proposed a genetic algorithm (GA) approach for seeding the k-means clustering method with
centers using a unique crossover operator that swaps adjacent centers. Nguyen and Cios [14] developed a
clustering method, called genetic algorithm k-means logarithmic regression expectation maximization
(GAKREM), capable of performing clustering without requiring the predetermined number of groups. It is a
hybrid of genetic algorithm, K-means, logarithmic regression, and expectation-maximization. Armano and
Farmani [15] suggested an approach the combination of k-means and artificial bee colony algorithms
(kABC), that used the artificial bee colony (ABC) algorithm to enhance k-means' capacity to identify global
optimal clusters in nonlinear partitional clustering situations. Karaboga and Ozturk [16] applied the ABC
algorithm as data clustering on classification benchmark problems. Zhang et al. [17] presented how to divide
N items into k clusters using the ABC algorithm. Zou et al. [18] introduced the cooperative ABC (CABC), an
expanded ABC method that outperforms the original ABC in addressing complex optimization problems. It
was used to solve issues with clustering. Bonab et al. [19] utilized the k-means algorithm with the ABC
algorithm and the differential evolutionary algorithm to get the optimal solution of objects in datasets and
images. Wang and Wang [20] employed a hybrid algorithm from the k-means and ABC algorithms to do the
clustering procedure. The present work is a step forward in this respect. We will address the defects of the
k-means algorithm using the proposed first ABC algorithm with a variable-length food source to predict the
optimal number of clusters (the value of k) and the initial centroids for the k-means method.
In the remaining portion of this paper, the k-means algorithm and ABC algorithm are briefly
reviewed in section 2. Section 3 describes the proposed method. Results are explained in section 4. Section 5
discusses the current findings. Section 6 presents the conclusion.
2. BACKGROUND
2.1. K-means clustering algorithm
K-means clustering algorithm is simple, an iterative, numerical, unsupervised, and non-deterministic
technique [21]. It is commonly used in data mining for grouping huge sets of datasets. It is a partitioning
clustering algorithm that divides the supplied datasets into k distinct clusters over an iterative operation that
meets a local minimum.
K-means algorithm starts with k, a user-specified parameter, initial cluster centers randomly chosen
from the dataset. In each iteration, each point at a given dataset is assigned to the closest cluster center. After
categorizing all data points into clusters, the new centroid of each cluster is re-calculated as the mean of all
cluster points. The procedure is continued until the centroids are still the same.
2.2. Artificial bee colony algorithm
The ABC algorithm is based on actual honey bee behavior [22]. It comprises three different groups
of bees: bees that are working, watching, and scouting. There are observers in the first half of the colony,
which are then employed artificial bees, and in the second half, which are the onlookers, there are employed
artificial bees. Only one bee is used in the collection of each type of food. This statement can be summarized
as saying that the number of employed bees equals the number of food sources. When the bees have depleted
the food supply, the hired bee takes on the role of a scout.
The ABC algorithm is divided into four stages: initialization, employed bees, spectator bees, and
scout bees, as presented in the following lines.
− Initialization: a random number of solutions is placed in the search space using (1):
𝑋𝑖𝑗 = 𝐿𝑏𝑗 + 𝑟𝑎𝑛𝑑 (0,1) ∗ (𝑈𝑏𝑗 − 𝐿𝑏𝑗) (1)
where Xi=1..., SN={xi1, xi2, …, xiD}, SN is the set of all possible solutions, D is the number of
optimization parameters that need to be determined, and the lower and upper limits of the solution site in
dimension j are defined as Lbj and Ubj, respectively.
− Employed bees phase: employed bees and candidate solutions have one-to-one communication.
Employed bees visit each of their solution candidates and change one dimension of the visited solution
until they find the solution using (2):

Int J Elec & Comp Eng ISSN: 2088-8708 
Optimal k-means clustering using artificial bee colony algorithm with … (Sabreen Fawzi Raheem)
5437
𝑉𝑖𝑗 = 𝑋𝑖𝑗 + 𝜑𝑖𝑗 ∗ ( 𝑋𝑖𝑗 − 𝑋𝑘𝑗) (2)
where j is a uniformly chosen dimension at random from {1,2, …, D}, φ is a numeric value generated at
random [-1 1], Xk solution is chosen at random from the population with the condition (i≠k). In a manner
similar to how the solutions in the bee's memory replace each other, when Vij is larger than the previous
position, the new solution replaces the old one; otherwise, the previous solution's location is retained.
Once all worker bees have completed their search, they perform a waggle dance to alert spectator bees to
their food sources.
− Onlooker bees phase: employed bees assess the fitness, fitnessi, of a solution Xi using (3):
𝑓𝑖𝑡𝑛𝑒𝑠𝑠𝑖 = {
1
1+𝑓(𝑋𝑖)
𝑖𝑓 𝑓(𝑋𝑖) ≥ 0,
1 + 𝑎𝑏𝑠 (𝑋𝑖) 𝑖𝑓 𝑓(𝑋𝑖) < 0.
(3)
Each onlooker bee visits a solution based on the selection probability of a candidate solution Xi using (4).
𝑃𝑖 =
𝑓𝑖𝑡𝑛𝑒𝑠𝑠𝑖
∑ 𝑓𝑖𝑡𝑛𝑒𝑠𝑠𝑖
𝑆𝑁
𝑖=1
(4)
When an onlooker bee chooses a solution, it looks for a new, better option. As a result, the employed bees
step controls the ABC algorithm's diversification behavior, whereas the observer bees step controls the
algorithm's intensification behavior.
− Scout bees step: in this phase, every food source that does not improve beyond a specific ‘limit’ of trials
is abandoned and replaced with a new location, which serves as a scout for the bee employed, using (1).
3. THE PROPOSED METHOD
To overcome the previously mentioned limitations of the k-means algorithm, this section discusses
the first mechanism to enhance the basic ABC algorithm to deal with variable-length representation
(ABCVL). Different lengths, i.e., numbers of features, individuals may focus on various search space areas.
Based on this ability, we first develop a new initialization technique to generate variable-length individuals to
provide a suitable diversity level for the whole. In addition, a special search equation is designed to deal with
variable-length individuals. The ABCVL algorithm was adopted to select the best number of clusters (K) and
the initial centroids for the k-means algorithm, as discussed below.
3.1. Initial population
Each food source is represented by a variable length from one source to another. We follow the
principle of variable-length chromosomes [23], [24] to represent individuals in the population. In this phase,
the population of SN individuals is randomly initialized with a different value of optimization parameters (k).
Each individual, ind, in the population is represented in the following:
𝑖𝑛𝑑𝑗 = {𝑠1,𝑠2,𝑠3, ⋯ 𝑠𝑘𝑗
} , 1 ≤ 𝑗 ≤ 𝑆𝑁
where SN is the number of food sources (solutions) in the population, the value of s is selected randomly
from the dataset, kj indicates that the individual j has k centers of clustering whose value varies from one
individual to another depending on the minimum, nCmin, and maximum, nCmax, number of centers, where
𝑛𝐶𝑚𝑖𝑛 ≤ 𝑘 ≤ 𝑛𝐶𝑚𝑎𝑥
3.2. Objective function
To measure the overall k-means clustering quality of food sources, we used different objective
functions (the nectar) based on the most important metrics to evaluate the goodness of a clustering, like
davies-bouldin index (DBI) measure [25], Silhouette coefficient, homogeneity measure, completeness
measure, v-measure, and Inertia. In addition to hybrid metrics, such as DBI-homogeneity-completeness
measure, DBI-Silhouette measure, and DBI-V-measure, are applied. Below is an explanation of the concept
of each of these metrics.
− The DBI measure is calculated by comparing each cluster's average similarity to the most similar. The
smaller the average similarity, the more distinct the groups are and the more accurate the clustering result.

 ISSN: 2088-8708
5438
− The Silhouette coefficient, aka the Silhouette score, is used to calculate the quality of a clustering method.
The result of this coefficient is in the range [-1,1], where: i) 1: means that clusters are well-defined and
distinct; ii) 0: means that clusters are indifferent to one another or that the distance between them is
insignificant; and iii) -1: means that clusters were misallocated.
− A homogeneity measure is a statistic that indicates the proportion of samples from a particular class that
belong to a single cluster. The fewer distinct classifications that are included inside a cluster are, the
better. The lower and upper limits of this measure should be 0.0 and 1.0, respectively, where higher is
better.
− Completeness measure: a perfectly complete clustering is one where all data points belonging to the same
class are clustered into the same cluster. Completeness and homogeneity are symmetrical.
− V-measure is equal to the harmonic mean of completeness and homogeneity.
− Inertia is a simple criterion derived from the sum of square error, which is used as the criterion clustering
function. It is the sum of distance squares between all samples and their centers of clustering.
− Hybrid scale 1: In this scale, the DBI measure is combined with homogeneity and completeness measures
using (5).
𝐻𝑦𝑏𝑟𝑖𝑑 𝑠𝑐𝑎𝑙𝑒1 = 0.5 ∗ 𝐷𝐵𝐼𝑠𝑐𝑜𝑟𝑒 + 0.25 ∗ homogeneityscore + 0.25 ∗ completenessscore (5)
− Hybrid scale 2: In this scale, the DBI measure is combined with Silhouette coefficient using (6).
𝐻𝑦𝑏𝑟𝑖𝑑 𝑠𝑐𝑎𝑙𝑒2 = 0.6 ∗ 𝐷𝐵𝐼𝑠𝑐𝑜𝑟𝑒 + 0.4 ∗ ( 1 − Silhouettescore) (6)
− Hybrid scale 3: In this scale, the DBI measure is combined with V-measure using (7).
𝐻𝑦𝑏𝑟𝑖𝑑 𝑠𝑐𝑎𝑙𝑒3 = 0.6 ∗ 𝐷𝐵𝐼𝑠𝑐𝑜𝑟𝑒 + 0.4 ∗ ( 1 − Vmeasure) (7)
3.3. Modified search equation
The basic ABC algorithm was improved to process variable-length solutions. We have modified the
search equation, namely MSEq, by following various modification directions, as discussed below. Figure 1
shows the possible cases to produce a new individual Vi from two individuals Xi and Xk, with the lengths L1
and L2, respectively.
3.3.1. MSEq1
In the first case, MSEq1, the modification position j is chosen within the range of the smaller
individual length that is shared between the two individuals, and (2) is applied. The length of the new
individual is the same as the length of the first individual. Figure 1(a) shows the schematic diagram of the
MSEq1 example, assuming that the length of the selected food sources (solutions) is 7 and 5, respectively.
The position j value, therefore, must be randomly chosen within the range [1,5]. In Figure 1(a), for instance,
the value of j is selected as 3.
3.3.2. MSEq2
In the second case, MSEq2, if L1 is less than L2 and the modification position j is chosen within the
range of L2 but outside L1, then the values of Xi are saved to Vi and the value of Xk,j is saved to Vi,j+1.
Therefore, the length of the new individual is L1+1. Figure 1(b) shows the schematic diagram of the MSEq2
example, assuming that the length of the selected food sources (solutions) is 3 and 5, respectively. The
position j value, therefore, is randomly chosen within the range [1,5]. If j falls outside 3, here is 4, then the
value of Xk,4 will be saved to Vi,4.
3.3.3. MSEq3
In the third case, MSEq3, if L1 is bigger than L2 and the modification position j is chosen outside L2,
(2) is applied on the value of Xi,j with any a random position at Xk,j. The rest positions at Xi are neglected.
Therefore, the length of the new individual is j.
Figure 1(c) shows the schematic diagram of the MSEq3 example, assuming that the length of the
selected food sources (solutions) is 5 and 3, respectively. The position j value, therefore, is randomly chosen
within the range [1,5]. If j falls outside 3, here is 4, then (2) is applied between Xi,4 and a random position at
Xk, here 2, to represent Vi,4. The length of the new individual is 4. As can be seen above, the MSEq1 differs
from MSEq2 and MSEq3 in the length of the new solution, which may vary from the original solutions at the
MSEq2 and MSEq3 but at MSEq1 still is the same as the first individual. The pseudocode of the ABCVL
algorithm is given in algorithm 1.

5439
(a) (b)
(c)
Figure 1. The schematic diagram of (a) MSEq1, (b) MSEq2, and (c) MSEq3
Algorithm 1. The pseudocode of the ABCVL algorithm
L1 is the length of Xi
L2 is the length of Xk
− Initialization phase using variable-length individuals (see Section 3.1);
− Repeat
− Employed bees phase using the MSEq technique;
− Onlooker bees phase using the MSEq technique;
− Scout bees phase;
− Memorize the best solution achieved so far;
− Until (termination conditions are met).
4. RESULTS
4.1. Tested datasets
To evaluate the efficiency of the ABCVL algorithm, three real-life datasets are used. The summary
details of these datasets are: i) Mall_Customers dataset [26] consists of 200 samples with five features:
Customer_ID (Unique ID for each customer), gender (Customer’s gender, male or female), age (Customer's
age), annual income (Customer's income per year, in $), and spending score (Customer’s spending score,
from 1-100). This dataset is to be tested for two features (annual income and spending score). The optimum
number of clusters is 5; ii) digits dataset [27] consists of 1797 8×8 grayscale images of handwritten digits.
Each feature is an integer in the range 0...16. The optimum number of classes is 10 (code 0...9); and
iii) breast cancer dataset [28] consists of 569 samples with two features (real or positive). It is divided into
two classes: 212 malignant and 357 benign samples with a dimensionality of 30. The optimum number of
classes is 2.
4.2. Parameters setting
In the artificial bee colony with variable-length (ABCVL) algorithm, the colony size (NP) was 40,
with the number of food sources (SN) equal to NP/2. The minimum (nCmin) and maximum (nCmax) number of
optimization parameters (D) were 2 and 15, respectively. The value of abandonment limit counter “limit”
was selected as SN*D. The maximum number of cycles was taken as 100.
4.3. Results
We carried a test using the ABCVL algorithm on two types of experiments. The first experiment,
Experiment I, was performed to evaluate the effectiveness of the objective function on the performance of the

 ISSN: 2088-8708
5440
ABCVL algorithm. On the other hand, the second experiment, Experiment II, was performed to evaluate the
variable-length individuals' effectiveness.
4.3.1. Experiment I
To select a suitable metric as an objective function, we tested eight metrics (Section 3.2) to evaluate
the ABCVL algorithm for clustering problems on three datasets with taking into consideration an evaluation
of performance. We executed the ABCVL algorithm ten times with each metric and chose the best one
among them. Table 1 shows the results of this experiment in terms of the best value, mean, optimal k, and the
mean of ten values of k, where the best results are highlighted in bold.
Table 1. The results of Experiment I
Datasets
Method
Breast_cancer
Digits
Mall_Customers
Mean
(k)
Optimal
k
Mean
Best
Mean
(k)
Optimal
k
Mean
Best
Mean
(k)
Optimal
k
Mean
Best
2.2
2
0.506289
0.504404
10.3
10
1.704200
1.690358
5.8
5
0.570686
0.566707
DBI
2.0
2
0.302735
0.302735
11.9
12
0.809863
0.809094
5.0
5
0.446068
0.446068
Silhouette
8.8
9
0.336008
0.334320
13.8
14
0.1611149
0.154497
-
-
-
-
Homogeneity
2.0
2
0.483191
0.483191
2.8
2
0.158477
0.152885
-
-
-
-
Completeness
3.0
3
0.529086
0.529086
12.1
12
0.201859
0.193189
-
-
-
-
V_measure
9.0
9
9601898
9432896.69
13.8
14
1052333
1043333.01
14.0
14
13067.55
12738.31
Inertia
2.0
2
0.517427
0.517427
11.5
10
0.986869
0.975439
-
-
-
-
Hybrid-scale1
2.0
2
0.423736
0.423736
.0
10
10
1.355614
1.337068
5.0
5
0.521398
0.521398
Hybrid-scale2
2.0
2
1.132735
1.120295
11.0
10
1.176221
1.164177
-
-
-
-
Hybrid-scale3
4.3.2. Experiment II
To test the performance of the ABCVL algorithm on clustering problems, we compared its results
with the basic k-means algorithm on three clustering datasets. The results of the current experiment, in terms
of the best value, mean and standard deviation produced by each algorithm throughout 30 runs, are shown in
Table 2, where the best results are highlighted in bold. In addition, the ABCVL algorithm with the hybrid-
scale2 function evolution curve is shown in Figure 2(a) to (c) for the three tested datasets. The coordinates of
the cluster center produced are given in Tables 3-5 for three tested datasets.
Table 2. The results of Experiment II
Dataset The basic K-means algorithm The ABCVL algorithm
Best Mean SD Best Mean SD
Mall_Customers 0.5213980 0.5895242 0.1056726 0.52139798 0.5209177 0.002586571
Digits 1.382326 1.503292 0.06446432 1.336862 1.345410 0.008708451
Breast_cancer 0.4237363 0.4237363 1.110×10-16
0.42373629 0.4237363 1.110×10-16
Table 3. Coordinates for cluster centers obtained by the ABCVL algorithm for the Mall_Customers dataset
Dimensionality=2 Clustering Centers for Mall_Customers dataset
1 2 3 4 5
Annual Income 73.0 54.0 30.0 99.80159093758257 54.436110937843466
Spending Score 88.0 48.0 73.0 81.04616983940875 99.0
Table 4. Coordinates for cluster centers obtained by the ABCVL algorithm for digits dataset
Dimensionality=64 Clustering centers for Digits dataset
1 2 3 4 5 6 7 8 9 10
0 0 0 0 0 0 0 0 0 0 0
1 3.01693 0 0.22672 0 0 0 0 0 0 0
2 8.35953 0 3.3012 13.1017 1 6 0 0 12 0
3 16.000 13.215 14.478 14.369 8.984 15 8 8 16 4
          
60 12.508 14.633 16.000 0.000 9.853 5 7 14 0 16
61 16.000 16.000 7.505 0 3.303 0 0 16 0 6
62 1.031 11.529 1.6 0 0 0 0 4 0 0
63 0 0 0 0 0 0 0 0 0 0

5441
Table 5. Coordinates for cluster centers obtained by the ABCVL algorithm for Breast_cancer dataset
Dimensionality=30 Clustering centers for Breast_cancer dataset
1 2
1 14.95 17.20
2 18.77 24.52
3 97.84 114.20
… 689.5 929.4
26 … …
27 0.2500 0.6566
28 0.08405 0.18990
29 0.2852 0.3313
0 0.09218 0.13390
5. DISCUSSION
The results in Table 2 indicate that the best metric for evaluating the clustering process is the
hybrid-scale2, i.e., the combination of DBI measure with the Silhouette coefficient. This metric gives the
optimal number of k for all datasets and all experiment runs. Therefore, this metric will be used in the current
system to evaluate the performance of the clustering method. The results have shown that applying the
current work produces better performance than setting the parameters using trial-and-error. In addition, using
the current work produces better results than the basic k-means algorithm relating to the best value, mean,
standard deviation, and speed of finding the best value for the same tested datasets. As shown in Figure 2, the
ABCVL gives a faster convergence speed for the Mall_Customers dataset in Figure 2(a) than it is with the
Breast_cancer dataset in Figure 2(c), which in turn is faster than it is with the digits dataset in Figure 2(b).
(a) (b)
(c)
Figure 2. The curve of the hybrid scale2 evolution for three tested datasets (a) Mall_Customers dataset,
(b) digits dataset, and (c) Breast_cancer dataset
6. CONCLUSION
Choosing a suitable combination of parameters for the k-means algorithm, like the number of
clusters k and the initial centroids, is a challenge and becomes more complicated when trading with complex
problems. Selecting reasonable parameters depends on the nature of problems and applications. One potential
solution to overcome this challenge using trial and error based on an expert valuation, but it may not be

 ISSN: 2088-8708
5442
performed efficiently in areas where the level of expertise is low. Moreover, even if expertise level is
relevant, specifying a fair value for each parameter can become a difficult task and time-consuming. Another
solution to determine the optimal number of clusters is semi-automatic approaches such as elbow, silhouette,
and gap statistics. Unfortunately, these approaches are subjective. An alternative solution is to assign each
parameter automatically. To achieve this goal, we proposed the first ABC algorithm with variable-length
individuals, ABCVL; in fact, we have found no such attempt so far for the ABC algorithm. In the ABCVL
algorithm, we first develop a new initialization technique to generate variable-length food sources to provide
a suitable diversity level for the whole. Then, a special search equation is designed to deals with variable-
length individuals. Overall, the current work achieved more encouraging findings than the traditional
k-means algorithm on three real-life clustering datasets. They seem to be consistent with those of other
studies and suggest that further work in this direction would be very worthwhile. In future work, we will take
the order of the data into account since it impacts the final results.
REFERENCES
[1] I. S. A. AL-Forati, A. Rashid, and A. Al-Ibadi, “IR sensors array for robots localization using k-means clustering algorithm,”
International Journal of Simulation: Systems, Science and Technology, Mar. 2019, doi: 10.5013/IJSSST.a.20.S1.12.
[2] K. Lee, H. W. Kim, C. Moon, and Y. Nam, “Analysis of vocal disorders using Cobweb clustering,” in 1st International
Conference on Artificial Intelligence in Information and Communication, Feb. 2019, pp. 120–123, doi:
10.1109/ICAIIC.2019.8669011.
[3] L. H. Trang, N. Le Hoang, and T. K. Dang, “A farthest first traversal based sampling algorithm for k-clustering,” in Proceedings
of the 2020 14th International Conference on Ubiquitous Information Management and Communication, Jan. 2020, doi:
10.1109/IMCOM48794.2020.9001738.
[4] S. van Steenkiste, M. Chang, K. Greff, and J. Schmidhuber, “Relational neural expectation maximization: unsupervised discovery
of objects and their interactions,” 6th International Conference on Learning Representations, ICLR 2018 - Conference Track
Proceedings, Feb. 2018
[5] A. Jabbar, “Local and global outlier detection algorithms in unsupervised approach: a review,” Iraqi Journal for Electrical and
Electronic Engineering, vol. 17, no. 1, pp. 1–12, Mar. 2021, doi: 10.37917/ijeee.17.1.9.
[6] S. Al-Dabooni and D. Wunsch, “Model order reduction based on agglomerative hierarchical clusterin,” IEEE Transactions on
Neural Networks and Learning Systems, vol. 30, no. 6, pp. 1881–1895, Jun. 2019, doi: 10.1109/TNNLS.2018.2873196.
[7] U. Maulik and S. Bandyopadhyay, “Genetic algorithm-based clustering technique,” Pattern Recognition, vol. 33, no. 9,
pp. 1455–1465, Sep. 2000, doi: 10.1016/S0031-3203(99)00137-5.
[8] D. W. van der Merwe and A. P. Engelbrecht, “Data clustering using particle swarm optimization,” in The 2003 Congress on
Evolutionary Computation, 2003, 2003, vol. 1, pp. 215–220, doi: 10.1109/CEC.2003.1299577.
[9] M. Omran, A. P. Engelbrecht, and A. Salman, “Particle swarm optimization method for image clustering,” International Journal
of Pattern Recognition and Artificial Intelligence, vol. 19, no. 3, pp. 297–321, May 2005, doi: 10.1142/S0218001405004083.
[10] M. Zhao, H. Tang, J. Guo, and Y. Sun, “Data clustering using particle swarm optimization,” in Lecture Notes in Electrical
Engineering, vol. 309, Springer Berlin Heidelberg, 2014, pp. 607–612., doi: 10.1007/978-3-642-55038-6_95.
[11] Y.-T. Kao, E. Zahara, and I.-W. Kao, “A hybridized approach to data clustering,” Expert Systems with Applications, vol. 34,
no. 3, pp. 1754–1762, Apr. 2008, doi: 10.1016/j.eswa.2007.01.028.
[12] T. Niknam and B. Amiri, “An efficient hybrid approach based on PSO, ACO and k-means for cluster analysis,” Applied Soft
Computing, vol. 10, no. 1, pp. 183–197, Jan. 2010, doi: 10.1016/j.asoc.2009.07.001.
[13] M. Laszlo and S. Mukherjee, “A genetic algorithm that exchanges neighboring centers for k-means clustering,” Pattern
Recognition Letters, vol. 28, no. 16, pp. 2359–2366, Dec. 2007, doi: 10.1016/j.patrec.2007.08.006.
[14] C. D. Nguyen and K. J. Cios, “GAKREM: a novel hybrid clustering algorithm,” Information Sciences, vol. 178, no. 22,
pp. 4205–4227, Nov. 2008, doi: 10.1016/j.ins.2008.07.016.
[15] G. Armano and M. R. Farmani, “Clustering analysis with combination of artificial bee colony algorithm and k-means technique,”
International Journal of Computer Theory and Engineering, vol. 6, no. 2, pp. 141–145, 2014, doi: 10.7763/IJCTE.2014.V6.852.
[16] D. Karaboga and C. Ozturk, “A novel clustering approach: artificial bee colony (ABC) algorithm,” Applied Soft Computing,
vol. 11, no. 1, pp. 652–657, Jan. 2011, doi: 10.1016/j.asoc.2009.12.025.
[17] C. Zhang, D. Ouyang, and J. Ning, “An artificial bee colony approach for clustering,” Expert Systems with Applications, vol. 37,
no. 7, pp. 4761–4767, Jul. 2010, doi: 10.1016/j.eswa.2009.11.003.
[18] W. Zou, Y. Zhu, H. Chen, and X. Sui, “A clustering approach using cooperative artificial bee colony algorithm,” Discrete
Dynamics in Nature and Society, vol. 2010, pp. 1–16, 2010, doi: 10.1155/2010/459796.
[19] M. B. Bonab, S. Z. M. Hashim, A. K. Z. Alsaedi, and U. R. Hashim, “Modified k-means combined with artificial bee colony
algorithm and differential evolution for color image segmentation,” in Advances in Intelligent Systems and Computing, vol. 331,
Springer International Publishing, 2015, pp. 221–231, doi: 10.1007/978-3-319-13153-5_22.
[20] X. M. Wang and J. B. Wang, “Improved artificial bee colony clustering algorithm based on k-means,” Applied Mechanics and
Materials, vol. 556–562, pp. 3852–3855, May 2014, doi: 10.4028/www.scientific.net/AMM.556-562.3852.
[21] K. P. Sinaga and M.-S. Yang, “Unsupervised k-means clustering algorithm,” IEEE Access, vol. 8, pp. 80716–80727, 2020, doi:
10.1109/ACCESS.2020.2988796.
[22] D. Karaboga, “An idea based on honey bee swarm for numerical optimization,” Tech. Rep., 2005.
[23] U. Maulik and S. Bandyopadhyay, “Fuzzy partitioning using a real-coded variable-length genetic algorithm for pixel
classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 41, no. 5, pp. 1075–1081, May 2003, doi:
10.1109/TGRS.2003.810924.
[24] Z. Qiongbing and D. Lixin, “A new crossover mechanism for genetic algorithms with variable-length chromosomes for path
optimization problems,” Expert Systems with Applications, vol. 60, pp. 183–189, Oct. 2016, doi: 10.1016/j.eswa.2016.04.005.
[25] S. Petrovic, “A comparison between the silhouette index and the davies-bouldin index in labelling IDS Clusters,” in 11th Nordic
Workshop on Secure IT-systems, 2006, pp. 53–64.
[26] Shwetabh123, “Mall_Customers,” Kaggle. https://www.kaggle.com/shwetabh123/mall-customers (accessed Jul. 15, 2021).

5443
[27] “sklearn.datasets.load_digits,” Scikit learn. Accessed: Jul. 15, 2021. [Online]. Available: https://scikit-
learn.org/stable/modules/generated/sklearn.datasets.load_digits.html#sklearn.datasets.load_digits
[28] “sklearn.datasets.load_breast_cancer,” Scikir learn. Accessed: Jul. 15, 2021. [Online]. Available: https://scikit-
learn.org/stable/modules/generated/sklearn.datasets.load_breast_cancer.html
BIOGRAPHIES OF AUTHORS
Sabreen Fawzi Raheem received the B.Sc. degree in Computer Science from
College of Science, University of Basrah, Iraq, in 2016. Currently, she is a master student at
the Department of Computer Science, College of Computer Science and Information
Technology, University of Basrah. She has authored or coauthored more than three refereed
journal and conference papers. Her research interests include artificial intelligence,
machine learning, and computational linguistic. She can be contacted at email:
sabreen.fawzi@uobasrah.edu.iq.
Maytham Alabbas is currently a Professor in the Department of Computer
Science at the University of Basrah where he has been a faculty member since 2003. He
received his Ph.D. degree (2013) in Computer Science from the University of Manchester,
UK. His Ph.D. thesis has been awarded the 2014 Best Thesis Prize of the School of Computer
Science at the University of Manchester. He received his M.Sc. (2002) and B.Sc. (1999) in
Computer Science from the University of Basrah, Iraq. His current research concerns are
artificial intelligence, NLP, machine learning, computational linguistic, and language
engineering. He has served on different conference and workshop program committees such
as IntelliSys 2021-2019, AMLTA 2019, ACLing 2018, and AISI 2018. He published more
than 22 journal papers and 16 conference papers. He has ACM professional membership. He
can be contacted at email: ma@uobasrah.edu.iq.

Optimal k-means clustering using artificial bee colony algorithm with variable food sources length

More Related Content

Similar to Optimal k-means clustering using artificial bee colony algorithm with variable food sources length (20)

More from IJECEIAES (20)

Recently uploaded (20)

Optimal k-means clustering using artificial bee colony algorithm with variable food sources length