The suitable distance function for fuzzy C-

Means clustering
Cite as: AIP Conference Proceedings 2578, 060006 (2022); https://doi.org/10.1063/5.0106185
Published Online: 03 November 2022

Joko Eliyanto and Sugiyarto Surono

ARTICLES YOU MAY BE INTERESTED IN

Proposing IoT and big data implementation model for business performance
AIP Conference Proceedings 2578, 050002 (2022); https://doi.org/10.1063/5.0106211

Logarithmic transformation for enhancing keypoint matching of SIFT in augmented reality

AIP Conference Proceedings 2578, 060010 (2022); https://doi.org/10.1063/5.0106221

Comparison of random forest algorithm, support vector machine and neural network for
classification of student satisfaction towards higher education services
AIP Conference Proceedings 2578, 060003 (2022); https://doi.org/10.1063/5.0106201

AIP Conference Proceedings 2578, 060006 (2022); https://doi.org/10.1063/5.0106185 2578, 060006

© 2022 Author(s).
 The Suitable Distance Function for
Fuzzy C-Means Clustering
Joko Eliyanto and Sugiyarto Surono a)

Department of Mathematics, Ahmad Dahlan University, Yogyakarta, Indonesia

a)
Corresponding author: [email protected]

Abstract. Fuzzy C-Means clustering is a form of clustering based on distance which apply the concept of fuzzy logic. The
clustering process works simultaneously with the iteration process to minimize the objective function. This objective
function is the summation from the multiplication of the distance between the data coordinates to the nearest cluster centroid
with the degree of which the data belong to the cluster itself. Based on the objective function equation, the value of the
objective function will decrease by increasing the number of iteration process. This research provide how we choose the
suitable distance for Fuzzy C-Means clustering. The right distance will meet the optimization problem in the Fuzzy C-
Means Clustering method and produce good cluster quality. They are Euclidean, Average, Manhattan, Chebisev,
Minkowski, Minkowski-Chebisev, and Canberra distance. We use five UCI Machine Learning dataset and two random
datasets. We use the Lagrange multiplier method for the optimization of this method. The result quality of the cluster
measure by their accuracy, Davies Bouldin Index, purity, and adjusted rand index. The experiment shows that the Canbera
distances are the best distances which provide the optimum result by producing minimum objective function 378.185. The
suitable distance for the application of the Fuzzy C-Means Clustering method are Euclidean distance, Average distance,
Manhattan distance, Minkowski distance, Minkowski-Chebisev distance, and Canberra distance. These six distances
produce a numerical simulation that derives the objective function fairly constant. Meanwhile, the Chebisev distance shows
the movement of the value of the objective function that fluctuates, so it is not in accordance with the optimization problem
in the Fuzzy C Means Clustering method.

INTRODUCTIONS
The Fuzzy Clustering method is a clustering method based on the membership degree of the data [1]. This method
allows the membership degree of the data can be considered into certain classification. In Fuzzy Clustering, the method
which mostly used is Fuzzy C-Means [2]. The objective function of the Fuzzy C-Means is the multiplication value of
the membership degree of the data in certain cluster with the square value of the distance difference between the data
coordinate and cluster centroid [3]. There are several distances that can be used on Fuzzy C-Means clustering. This
distance is defined as a function that meet the required criteria [4]. First, the value of distance between two coordinates
is non-negative. Second, the value of the distance is zero if and only if there are two points in the same coordinate.
Third, the distance between A and B or B and A is equivalent. Fourth, the distance function has to meet the concept
of triangle inequality [5].

The research of the effect of the distance in clustering method has been done a lot. A few of the research conclude
with the conclusion where the various distance function that being used produces an output which are not much
different and there are no sign of the distance that appear dominant [6]–[8]. The effect of the variation of distance
function which applied to clustering method, Euclidean distance, Manhattan / City Block distance, Chebisev /
Maximum distance, and Minkowski distance, have already identified on K-Means Clustering algorithm [6], [9], [10].
In another research, Euclidean distance, Manhattan/City Block distance, Canberra distance, and Chebisev distance,

The 3rd International Conference on Engineering and Applied Sciences (The 3rd InCEAS) 2021
AIP Conf. Proc. 2578, 060006-1–060006-12; https://doi.org/10.1063/5.0106185
Published by AIP Publishing. 978-0-7354-4233-7/$30.00

060006-1
are applied and evaluated on fuzzy clustering algorithm. The result of that research concluded that the result of the
clustering process is very dependent on the dataset that being applied [11], [12].

Apart from the distance function that already mentioned, there are another variation of the distance function which
can be applied on clustering algorithm. There are Standardized Euclidean distance, Mahalanobis distance, Cosine
distance, Spearman distance, Canberra distance, Bray Curtiz distance, Average distance, Chord distance, Weighted
Euclidean distance, Hausdorf distance, and Minkowski-Chebisev distance [7], [13]–[16]. The different application of
a certain distance function can increase the performance quality of the clustering algorithm. For example, the
combination of the Minkowski and Chebisev distance is proven to improving the algorithm of the Fuzzy C-Means
Clustering [15]. The output result can be improved more by applying the reduction of PCA dimension on the high
dimension data [17], [18]. Another way to improve the performance quality of the clustering algorithm is by using
Average distance [19], [20].

The main objective from the Fuzzy C-Means clustering is to minimize the value of the objective function [21].
The objective function of Fuzzy C-Means clustering is a function that shows the error value. Through the optimization
process, the more the number of iterations, the lower the value of the objective function. This is the basis for
optimization of Fuzzy C-Means clustering. Intuitively, this method improving the result on each iteration by updating
the membership degree on every dataset point. The membership degree matrix represent the membership degree of a
certain data based on its distance to the existing centroid clusters [22]. This research focused on the observation to
determine the suitable distance to be applied on Fuzzy C-Means clustering. We use five UCI machine learning data,
which are Hill Valley dataset, Iris dataset, Seeds dataset, and Sonar dataset. We also provide two another random
dataset with certain specification functioned as an addition variable to analyse. We limit our research to the
development of the basic Fuzzy C-Means clustering method with the aforementioned seven distance functions.

METHODS

Fuzzy C-Means Clustering

Fuzzy C-Means (FCM) is a form of clustering method which enable one set of data to be classified into two or
more clusters [20]. This method is invented by Dunn in 1973 and improved further by Bezdek in 1981. This method
is commonly used to introduce certain pattern of data. This method is based on minimalization the following objective
function:

= − ,1 ≤ <∞
(1)

Where m is any real number valued more than 1, uij is a membership value of xi, on cluster j, xi shows the data in
i, cj is a cluster centroid in j, and ‖∗‖ is a norm which determine the distance between data and cluster centroid. Fuzzy
partition is determined by repeated optimization from the objective function which mentioned below with an update
of matrix membership uij and cluster centroid cj by:

1
=
− (2)
∑
‖ − ‖

∑ ∙
=
∑ (3)

060006-2
 ( )
The iteration process will stop when − ( ) < , where is a termination criteria between 0
and 1, where k is the number of iteration process. This procedure is convergent to local minimum value of Jm.

The Fuzzy C-Means Clustering algorithm is consist of the following steps [23]:

a. Initialization of = , ( )
( ) ( )
b. In steps of k : measure the vector of the cluster centroid = with

∑ ∙
=
∑
( ) ( )
c. Updating ,

1
=
−
∑
‖ − ‖
( ) ( )
d. If − < , then stops the iteration, if not, back to step 2.

Objective function and membership function are closely related in this method. We assume that every data
coordinate already has their own value of membership degree for each existing cluster on the initial process of
iteration. Then, this value will be updated over and over by using equation (2). During this process, equation (1) will
also be updated until reach its minimum value. Figure 1 (a) illustrate the value distribution of the data membership in
the initial of iteration process. Afterwards, the output of the iteration will be updated in every iteration (Figure 1 (b)).
The iteration process will be terminated when the value of objective function reached their minimum where each data
on every cluster has their ideal membership distribution which illustrated on Figure 1 (c). The ideal membership
distribution of the data is occurred when every data are gathered in their closest cluster centroids.

FIGURE 1. Membership Function Update During Iteration Process

Lagrange Multiplier

To maximize or minimize ( , ) against the constraint then solve the system of equations

∇ ( , )= ∇ ( , ) (4)

And

g( , ) = 0 (5)

060006-3
 For x, y and . Each point x, y is a critical point for the extreme value problem with the constraint and is called
the Lagrange multiplier.

Function ( , ) with constraint ( , ) = will reach the minimum or maximum by form it into a Lagrange
function which is defined as
L( , , ) = ( , ) + ( ( , ) − ) (6)

To get a solution to the optimization problem, the following necessary and sufficient conditions are required:

Necessary Condition:

= 0, = 0, =0 (7)

Sufficient Condition:

x For the maximization problem, then the main determinant of the Hessian matrix is negative definite.
x For the minimization problem, then the main determinant of the Hessian matrix is positive definite.

Distance Functions
Euclidean Distance

Euclidean distance is known as the most common and standard distance function to apply for clustering method.
This function for point x and y is define with this following equation [19]:

( , )= ( − ) (8)

Where and are the value of x and y on dimension n.

Manhattan Distance

Manhattan distance or City Block distance is defined as an addition of each its attributes. Hence, for two points of
data x and y on dimension n, the distance between them is define as [19]:

( , )= | − | (9)

Where and are the value of x and y on dimension n.

Chebisev Distance

Chebisev distance is also known as maximum distance. This distance is defined as the maximum value of distance
from its attributes that exist. The distance for two points of data x and y on dimension n is define as [19]:

( , )= | − | (10)

060006-4
 Minkowski Distance

Minkowski distance is a measurement from metric distance which define as [24]:

( , , )= | − | (11)

With p is a Minkowski parameter. On Euclidean distance ( = 2), Manhattan distance ( = 1), and Chebisev
distance ( → ∞). The metric condition for this function will satisfied as long as ≥ 1.

Minkowski-Chebisev Distance

This distance function is invented by Rodrigus (2018) and Novianty (2019) uses the combination of Minkowski-
Chebisev distance for Fuzzy C-Means Clustering. The combination of Minkowski and Chebisev distance with weight
w1 ¬and w2 are shown on equation (7). When w1 is greater than w2, this distance function will looks similar to
Manhattan distance and vice versa. The Minkowski-Chebisev distance is define as [25]:

( , , , , )= ( , , )+ ( , ) (12)

Canbera Distance

Canberra distance is an addition of the absolute ratio difference between two points of data [26]. The equation to
measure this distance function is [27]:

| − |
( , )= (13)
| |+| |

This distance is very sensitive to the alteration if the two points of data are close to 0. This distance is applied because
of its properties similarities with Manhattan distance.

Average Distance

Average distance is a modification of Euclidean distance. This modification is developed to improve the quality
of clustering output [24]. This distance is define with this following equation [19]:

1
( , )= ( − ) (14)

060006-5
 Datasets
This research based on five UCI Machine Learning dataset dan two random with three class. TABEL 1 illustrate
the detail information of the applied dataset.

TABLE 1. Datasets

Dataset Row Column Class

Random 1 150 4 3
Random 2 210 7 3
Hill-Valley 1484 8 10
Iris 208 60 2
Seeds 606 100 2
Sonar 100 2 2
Yeast 400 2 4

Cluster Evaluation
Cluster Evaluation has to be applied to find out the accuracy level of the clustering algorithm and to perceive the
existing classification. This research use 3 testing methods, Accuracy test, Purity test and Davies Bouldin Index, for
cluster evaluation [15][18].

Accuracy

Accuracy is calculated by adding up the number of objects that fall into the i-th cluster, where the correct one in
the original class is then divided by the number of data objects. Good accuracy results if all clusters match the original
class and then divided by the number of data will produce a value of 1.

Purity

Purity is used to calculate the purity of a cluster. A bad cluster has a purity value close to 0. This means that there
are no cluster results that match the original class, while a good cluster has a purity value of 1 . So, it can be interpreted
that the results of the cluster are in accordance with the original class.

Adjusted Rand Index

Adjusted Rand Index (ARI) is a Rand Index that has been fixed. ARI is used to measure cluster quality by
comparing the predicted class with the original class. The Rand Index value is in the interval , while the ARI value is
in the interval . The Rand Index value is in the interval [0,1] while the ARI value is in the interval [-1, 1] . The greater
the ARI value means the better the quality of a cluster.

Davies Bouldin Index

The Davies–Bouldin index (DBI), introduced by David L. Davies and Donald W. Bouldin in 1979, is a metric for
evaluating clustering algorithms. This is an internal evaluation scheme, where the validation of how well the clustering
has been done is made using quantities and features inherent to the dataset. This has a drawback that a good value
reported by this method does not imply the best information retrieval. Due to the way it is defined, as a function of
the ratio of the within cluster scatter, to the between cluster separation, a lower value will mean that the clustering is
better.

060006-6
 Research Method
This research is in the form of a numerical simulation. First we define Fuzzy C-Means Clustering with 7 different
distances namely Euclidean, Manhattan, Chebisev, Minkowski, Minkowski-Chebisev, Canbera, Average. This
method is an optimization problem. Then, we solve this problem using the Lagrange method. From the last step, we
obtain the Hessian matrix to see whether the optimization problem that is constructed is a maximization or
minimization problem. Then to verify it numerically, we implement Fuzzy C-Means Clustering in python
programming language and then record the results. FIGURE 2 is our research flowchart.

FIGURE 2. Research Flowchart

RESULT & DISCUSSION

First, we define the Fuzzy C-Means Clustering method with seven distance functions. Then we optimize using the
Lagrange Mutiplier method to obtain the Hessian matrix results for each distance function presented in TABLE 2.
The seven distances produce several possibilities for the Hessian Matrix, namely indefinite, semi-definite positive, or
positive definite. This means that there is a high probability that these functions are minimization problems. This result
will be verified by further numerical experimentation to see if the pattern of the minimization problem is seen in the
iteration process of the method.

060006-7
 TABLE 2. Hessian Matrix of Each Distance Function

Dataset Matrix Hessian

Euclidean indefinite or semi-definite positive or positive definite
Manhattan indefinite or semi-definite positive or positive definite
Minkowski indefinite or semi-definite positive or positive definite
Chebisev positive semi definite
Minkowski-Chebisev indefinite or semi-definite positive or positive definite
Canbera indefinite or semi-definite positive or positive definite
Average indefinite or semi-definite positive or positive definite

The Fuzzy C-Means Clustering method with 7 kinds of distance functions is implemented in a program written in
python. Parameter values which used on this research are : = 5, = 0.5, = 0.5, = 2, = 100, =
10 . The values of objective function is observed in every iteration process. The output comparison of the clustering
process on each dataset is done by normalizing the output value of the main objective by using this following equation
(15):

= × 100% (15)
( )

FIGURE 3 illustrate the result of the cluster evaluation average output in each of the distance which applied on
Fuzzy C-Means Clustering algorithm. The result shows that the clustering process output on each distance function is
not significantly different. This result is consistent with [6]. Nevertheless, the Euclidean distance and Average distance
are achieving a better result than the other distance function. The same result is also achieved on another research
[10].

Cluster Result Evaluation

1
0.8
0.6
0.4
0.2
0
Accuracy Purity ARI

Euclidean Manhattan Chebisev Minkowski Mikowski-Chebisev Canbera Average

FIGURE 3. Cluster Result Evaluation

Apart from the desired clustering output, the low value of objective function which should be minimum is also
the ideal result that we want to achieve. FIGURE 4 shows the average value of objective function value which achieved
on each distance function for 100 times iteration. The minimum values of objective function is achieved on 3 distance
functions; Euclidean distance, Manhattan distance, and Minkowski-Chebisev distance. This result concludes that the
combination of Minkowski distance and Chebisev distance able to improve the output of Fuzzy C-Means Clustering
algorithm. This result also mentioned on different research [15] with 0.546 as the minimum value. The value of
objective function on Manhattan distance is very different than other distance function which also shown in another
research [9].

060006-8
 Objective Function Value
Objective Functioon
355286.7454
3466.79081
814844.4742
1681268.962
102126.4825
34970252.07
3567161.56

0 5000000 10000000 15000000 20000000 25000000 30000000 35000000 40000000

Average Canbera Mikowski-Chebisev Minkowski Chebisev Manhattan Euclidean

FIGURE 4. Objective Function Value

The results of the clustering evaluation of the Fuzzy C-Means method are presented in TABLE 3. To determine
the best distance, scoring is performed which is shown in TABLE 4. The best distance obtained is the Canbera distance
with the highest score on all sides.
TABLE 3. Cluster Evaluation

Distance Function Iteration Time Computation Objective Function Accuracy DBI ARI Purity

Euclidean 40 5,468 694.389.760.824 0,611 1,146 0,307 0,641

Manhattan 72 12,854 67.247.168.149.060 0,624 1,577 0,308 0,657

Minkowski 58 16,399 154.255.495.471 0,606 1,827 0,308 0,659

Chebisev 93 12,284 11.921.246.033 0,368 1,532 0,190 0,577

Minkowski-Chebisev 63 19,923 59.104.443.850 0,481 1,548 0,276 0,642

Canbera 46 7,984 378.185 0,479 1,375 0,339 0,658

Average 58 15,712 6.943.897.619 0,601 1,854 0,309 0,660

TABLE 4. Scoring

Distance Function Iteration Time Computation Objective Function Accuracy DBI ARI Purity Total
Euclidean 7 7 2 6 7 3 2 34
Manhattan 2 4 1 7 3 4 4 25
Minkowski 5 2 3 5 2 5 6 28
Chebisev 1 5 5 1 5 1 1 19
Minkowski-Chebisev 3 1 4 3 4 2 3 20
Canbera 6 6 7 2 6 7 5 39
Average 4 3 6 4 1 6 7 31

060006-9
FIGURE 5 shows the comparison of objective function based on iteration number. The other distance function
parameter, the combination of Minkowski and Chebisev distance, does not produce the desired output, where the
tendency of the objective function values are increasing on Yeast dataset, Sonar dataset, Seeds dataset, and Random
2 dataset. The most undesired result occurs in the application of Chebisev distance where the tendency of the objective
function values are fluctuated. However, in different research, the usage of this distance function resulting on suitable
performance when another parameters are added [13], [27]. In fact, the performance from the application of Chebisev
distance still could be improved even further [9]. The results in this research are very dependent in the starting
initialization value of the matrix. That’s why the clustering method is also known as non-deterministic method.
Nevertheless, the output of this research is considered beneficial for further analysis on how the Fuzzy C-Means
Clustering method should be applied.

FIGURE 5. Objective Function Value Movement Based on Iteration

Good clustering results are not enough for the Fuzzy C-Means Clustering method. This method is a
minimization problem, so the value of the objective function will decrease continuously for each iteration. If this last
point is not met, then the modification of the distance function applied is said to be incorrect. For further research, it’s
essential to determine on how the termination criteria should be based on. The termination criteria in this research is
based on the concept where the significant value of clustering result is constant (below the values of H ). This criteria
is not exactly accurate, in fact, even when the minimum value of objective function is achieved, the objective function
valued below H are always continuously run the algorithm. On top of that, the possibility where the value of the
objective function in the next iteration process got worse could be occurred.

060006-10
 CONCLUSION
This research concluded that the most suitable distance functions to apply for Fuzzy C-Means clustering method
are Euclidean distance, Average distance, Manhattan distance, Minkowski distance, Minkowski-Chebisev distance,
and Canberra distance. Furthermore, this research shows that not every distance function will satisfy the minimization
problem even though the majority of the dataset that being applied are met. Based on the result of this research, it’s
important to visualize the tendency of the objective function values on Fuzzy C-Means clustering method. On the
other hand, focusing on the result of the clustering process is also important. It aims to achieving suitable result through
proper method. In the future, the research on the evaluation of the termination criteria and another different application
of distance functions could be analyzed further.

REFERENCES
1 M. Zhang, W. Zhang, H. Sicotte, and P. Yang, “A New Validity Measure for a Correlation-Based Fuzzy C-
Means Clustering Algorithm,” pp. 3865–3868, 2009.
2 N. Grover, “A study of various fuzzy clustering algorithms,” Int. J. Eng. Res., 2014.
3 J. Nayak, B. Naik, D. P. Kanungo, and H. S. Behera, “ENGINEERING PHYSICS AND MATHEMATICS A
hybrid elicit teaching learning based optimization with fuzzy c-means ( ETLBO-FCM ) algorithm for data
clustering,” AIN SHAMS Eng. J., 2016, doi: 10.1016/j.asej.2016.01.010.
4 J. Hernadi, Analisis Real Elementer dengan Ilustrasi Grafis & Elementer. Erlangga, 2015.
5 G. R. Bartle and R. Sherbet, Introduction to Real Analysis (4thed.). United States of America: Hamilton
Printing Company, 2010.
6 P. Grabusts, “The choice of metrics for clustering algorithms,” Proc. 8th Int. Sci. Pract. Conf., 2011.
7 V. Kumar, J. K. Chhabrea, and D. Kumar, “Performance evaluation of distance metrics in the clustering
algorithms,” INFOCOMP J. Comput. Sci., 2014.
8 Y. S. Thakare and S. B. Bagal, “Performance evaluation of K-means clustering algorithm with various
distance metrics,” Int. J. Comput. Appl., 2015.
9 O. A. M. Jafar and R. Sivakumar, “Hybrid Fuzzy Data Clustering Algorithm Using Different Distance
Metrics : A Comparative Study,” no. 6, pp. 241–248, 2014.
10 A. Singh, A. Rana, and U. Pradesh, “K-means with Three different Distance Metrics,” vol. 67, no. 10, pp. 13–
17, 2013.
11 B. Charulatha, P. Rodrigues, and T. Chitralekha, “A comparative study of different distance metrics that can
be used in Fuzzy Clustering Algorithms,” IJETTCS. Natl. Conf. Archit. Softw. Syst. Green Comput., 2013.
12 V. P. Mahatme and K. K. Bhoyar, “Impact Of Distance Metrics On The Performance Of K-Means And Fuzzy
C-Means Clustering-An Approach To Assess Student’s Performance In E-Learning Environment,” Int. J. Adv.
Res. Comput. Sci., 2018.
13 S. Boddana and H. Talla, “Performance Examination of Hard Clustering Algorithm with Distance Metrics,”
Int. J. Innov. Technol. Explor. Eng., vol. 9, no. 2S3, pp. 172–178, 2019, doi: 10.35940/ijitee.b1045.1292s319.
14 N. Gueorguieva, I. Valova, and G. Georgiev, “ScienceDirect ScienceDirect M & MFCM : Fuzzy C-means
Clustering with Mahalanobis and Minkowski Distance Metrics,” Procedia Comput. Sci., vol. 114, pp. 224–
233, 2017, doi: 10.1016/j.procs.2017.09.064.
15 P. Noviyanti, “Fuzzy c-Means Combination of Minkowski and Chebyshev Based for Categorical Data
Clustering,” Univ. Ahmad Dahlan, 2018.
16 A. S. Shirkhorshidi, S. Aghabozorgi, and T. Y. Wah, “A Comparison Study on Similarity and Dissimilarity
Measures in Clustering Continuous Data,” pp. 1–20, 2015, doi: 10.1371/journal.pone.0144059.
17 J. Eliyanto, Sugiyarto, Suparman, I. Djakaria, A. H. Mustafa, and Ruhama, “Dimension Reduction Using Core
and Reduct to Improve Fuzzy C-Means Clustering Performance,” Technol. Reports Kansai Univ., 2020.
18 S. Surono and R. D. A. Putri, “Optimization of fuzzy c-means clustering algorithm with combination of
minkowski and chebyshev distance using principal component analysis,” Int. J. Fuzzy Syst., 2020, doi:
10.1007/s40815-020-00997-5.
19 G. Gan, C. Ma, and J. Wu, “Data clustering: theory, algorithms, and applications,” Soc. Ind. Appl. Math.,
2020.
20 J. Han and M. Kamber, “Data Mining : Concepts and Techniques Third Edition,” Morgan Kaufmann.
Amsterdam, 2011.

060006-11
21 J. C. Dunn, “A Fuzzy Relative of the ISODATA Process and Its Use in Detecting Compact Well-Separated
Clusters,” J. Cybern., 1973.
22 J. C. Bezdek, Pattern recognition with fuzzy objective function algorithms. New York: Plenum Press, 1981.
23 A. K. T. Andu and A. S. Thanamani, Multidimensional Clustering Methods of Data Mining for Industrial
Applications. 2013.
24 T. Brunello, D. Bianchi, and E. Enrico, Introduction To Computational Neurobiology and Clustering.
Singapore: World Scientific Publishing, 2007.
25 O. Rodrigues, “Combining minkowski and cheyshev: new distance proposal and survey of distance metrics
using k-nearest neighbours classifier,” Pattern Recognit. Lett., vol. 110, pp. 66–71, 2018, doi:
10.1016/j.patrec.2018.03.021.

26 B. G. N. Lance and W. T. Williams, “Computer programs for hierarchical polythetic classification (" similarity
analyses "),” 1964.
27 C. Science, “Performance Evaluation of Distance Metrics in the Clustering Algorithms,” no. 1, pp. 38–51,
2014.

060006-12