You are on page 1of 20

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361805217

Hybrid binary whale with harris hawks for feature selection

Article  in  Neural Computing and Applications · July 2022


DOI: 10.1007/s00521-022-07522-9

CITATIONS READS

9 325

7 authors, including:

Ranya Wajih Said Jadid Abdulkadir


Universiti Teknologi PETRONAS Universiti Teknologi PETRONAS
3 PUBLICATIONS   59 CITATIONS    94 PUBLICATIONS   1,508 CITATIONS   

SEE PROFILE SEE PROFILE

Qasem Al-Tashi
University of Texas MD Anderson Cancer Center
39 PUBLICATIONS   1,039 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Metaheuristic Algorithms View project

Wrapper Multi-objective Feature Selection View project

All content following this page was uploaded by Said Jadid Abdulkadir on 08 July 2022.

The user has requested enhancement of the downloaded file.


Neural Computing and Applications
https://doi.org/10.1007/s00521-022-07522-9 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Hybrid binary whale with harris hawks for feature selection


Ranya Alwajih1 • Said Jadid Abdulkadir1 • Hitham Al Hussian1 • Norshakirah Aziz1 • Qasem Al-Tashi2,3 •

Seyedali Mirjalili4,5 • Alawi Alqushaibi1

Received: 4 October 2021 / Accepted: 7 June 2022


Ó The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2022

Abstract
A tremendous flow of big data has come from the growing use of digital technology and intelligent systems. This has
resulted in an increase in not just the dimensional issues that big data encounters, but also the number of challenges that big
data faces, including redundancies and useless features. As a result, feature selection is offered as a method for eliminating
unwanted characteristics. This study introduces the BWOAHHO memetic technique, which combines the binary hybrid
Whale Optimization Algorithm (WOA) with Harris Hawks Optimization (HHO). A transfer function to transfer continuous
characteristics to binary to fulfill the feature selection nature condition. The efficiency of the selected attributes is assessed
using a wrapper k-Nearest neighbor (KNN) Classifier. About 18 benchmark datasets obtained from UCI repository were
utilized to measure the proposed method’s proficiency. The performance of the novel hybrid technique was evaluated by
comparing to that of WOA, HHO, Particle Swarm Optimization (PSO), the Genetic Algorithm (GA), and the WOASAT-2.
With the new hybrid feature selection method, the WOA algorithm’s efficiency was improved. Classification accuracy,
average fitness, average selected attributes, and computational time were all used as performance indicators. In terms of
accuracy, the proposed BWOHHO algorithm compared with 5 similar metaheuristic algorithms. The BWOAHHO had a
classification accuracy of 92% in the 18 datasets, which was higher than BWOA (90%), BPSO (82%), and BGA (82%).
(83%), the fitness measures of the BWOHHO algorithm are 0.08, which is lower than the average fitness of BWOA, BPSO,
and BGA., in terms of selected attribute size compare the proposed BWOAHHO algorithm to the results obtained by the
other five techniques The average selected feature sizes for BWOAHHO, BWOA, BHHO, BGA, and WOASAT-2 were
18.07, 20.12, 22.3, 22.32, 22.40, and 15.99, respectively, and computing time for the proposed BWOHHO was 7.36 in
second which was the lowest computed value. To determine the significance of BWOAHHO, a statistical one-way
ANOVA test was used. When compared to existing algorithms, the proposed approach produced better results.

Keywords Machine learning  Classification  Feature (variable) selection  Whale optimization algorithm 
Hybrid feature algorithm  Harris hawks optimization

& Said Jadid Abdulkadir 1 Introduction


Saidjadid.a@utp.edu.my
1
In today’s era of machine learning, variable selection, is an
Centre for Research in Data Science (CeRDaS), Computer essential pre-processing step. Noise or irrelevant data,
Information Science Department, Universiti Teknologi
PETRONAS, SeriIskandar, Perak, Malaysia which can damage system acacuracy, is commonly asso-
2 ciated with the large expansion in the scale of gathered data
Department of Imaging Physics, The University of Texas MD
Anderson Cancer Center, Houston, TX, USA and the enormous volume of information [1, 2]. Feature
3 selection is a strategy for removing redundant and irrele-
Faculty of Administrative and Computer Sciences, University
of Albaydha, Albaydha, Yemen vant features that is frequently utilized in many domains.
4 Feature selection is essentially the key to determining
Centre for Artificial Intelligence Research and Optimization,
Torrens University Australia, Fortitude Valley, QLD 4006, feature subsets from the source datasets [3]. The curse of
Australia dimensionality, as described by Bellman in [4], The fun-
5
Yonsei Frontier Lab., Yonsei University, Seoul, Republic of damental issue with feature selection is that the quantity of
Korea characteristics increases system inaccuracy, where the

123
Neural Computing and Applications

quantity of attributes in a search space with n features is 2n methods [14]. In filter approaches, an independent criterion
[1, 5]. TAs a consequence, in order to increase system is applied, and the evaluation of the subset goodness does
accuracy and relevance, there is a great need to limit the not depend on any classifier or learning algorithm. Due to
quantity of redundant attributes. To locate the represent their simplicity, filter methods are used, and good results
features, attribute selection models go through four major have been reported for practical applications. However, the
stages [6]. Feature selection process begin with generating performance of the selected features is neglected by filter
subsets using searching algorithms, then evaluating the methods [15]. Wrapper techniques evaluate variable subset
produced attributes until the stopping requirement is met, performance using a black box. Wrapper methods tend to
finally, validate the selected features. yield better results, but the computing time required to
Attribute selection search strategies have changed over construct the feature subset is a major disadvantage [14]. In
time to achieve the optimal results. In the literature, con- an embedded approach, attribute selection and classifier
ventional search methods such as sequential search (SFS) learning are combined. [1]. Among the known EC tech-
and (SBS) were very popular in the past [6]. Layering, niques, only a few algorithms can accomplish embedded
stagnation in local optima, and expensive computing are feature selection [16, 17].
the shortcomings of such approaches [1]. To avoid the Despite the good results obtained by WOA, it has few
nesting effect of the prior approaches, floating search, internal settings and a limited number of heuristics [12].
sequential forward/backward floating search (SFFS) and The proposed method is a combination between the WOA
(SBFS) were proposed [7]. Nonetheless, as the search field and the HHO algorithms. Where exploration part is taken
expands, floating search techniques have a tendency to from HHO, and the exploitation part is carried from WOA.
miss the best solutions [8]. The proposed method aims to resolve this issue by
The evolution of feature selection led towards a more hybridizing the WOA and the HHO algorithms. The new
complex search algorithms known as metaheuristic algorithm is called BWOHHO, where its exploration part is
Approaches. For instance, there’s many efforts made to taken from HHO, and the exploitation part is carried from
boost the effectiveness of Swarm Intelligence (SI) , and WOA. The decision to combine WOA and HHO algo-
Evolutionary Algorithms (EA) [7]. GA is an EA algorithm, rithms was primarily driven by WOA’s simplicity and
while the algorithms that inspired from swarms’ behaviors effectiveness, as well as HHO’s stochastic search. The
such as Artificial Bee Colony (ABC) [8], Grey Wolf combination of these two heuristic methods is expected to
Optimizer (GWO) [9, 10], HHO [7], Ant Lion Optimiza- result in a simplified, functional hybrid algorithm.
tion (ALO) [11], and WOA [12] are gathered under SI The remaining content is arranged: Section 2 reviews
category. the recent studies of feature selection and a detail
WOA [12] is a recently developed optimization method description of both WOA and HHO. Section 3 clearly
by Mirjalili and Lewis, which was stimulated by the describes the proposed memetic algorithm. The experi-
behavior of humpback whale hunting strategies. To sur- mental setup of this study is explained in Sect. 4, Sect. 5
vive, humpback whales pursue and feed on small schools shows the results of the proposed hybrid algorithm and
of fish or krill near the surface. They use unique maneuvers compares them to state-of-the-art approaches. Finally, Sect.
to immobilize their victim, including as swimming in a 6 brings the paper to a conclusion.
circle or spiraling around them. This is known as a bubble-
net attack. The humpback whale hunts in two stages:
exploration and exploitation. The exploration interval 2 Related work
begins when the whales are randomly searching for prey
until the certain point the whales start exploiting the search In recent years, attraction in hybrid metaheuristics has
space by performing bubble-net hunting. grown significantly between evolutionary computation
HHO [7, 13] was designed to mimic the hunting style of researchers. A common paradigm for integrating the ben-
Harris hawks. The HHO exploration phase begins with the efits of both methods while reducing their limitations is
hawk setting a value to identify the perching site and then hybridization of two algorithms. The integrated strategies
observing the target. The hawks employed several diverse have been successful in resolving the defined problem and
hunting tactics depending on the victim’s energy and combining these strategies can improve the results obtained
capacity to flee when the exploitation phase began. Smooth individually. We can improve the discovery and exploita-
besiege, tough besiege, smooth besiege with gradual tion of algorithms by combining algorithms [18]. Accord-
speedy leap, and tough besiege with gradual speedy leap ing to Talbi [18], hybrid metaheuristics algorithms design
are the four foremost search strategies in this algorithm. categorized into Low-level which a specific metaheuristic
Each new subset must be evaluated using an evaluation function is altered by another metaheuristic, and High-level
function. Filter, wrapper, and embedded are evaluation which various metaheuristics are self-contained. In Low-

123
Neural Computing and Applications

level hybrid algorithms, there are two hybridization tech- objective of this study is to improve the dynamically
niques: Low-level (Rely and Teamwork) Hybrid (LRH), adjusting the search agents of exploitation to avoid local
(LTH). The LRH class of hybrid algorithm entails optima. Moreover, in [33], a new hybrid technique com-
embedding a particular metaheuristic within a single-so- bines the HHO algorithm with the cuckoo search algorithm
lution metaheuristic, whereas A metaheuristic is incorpo- (CS) and chaotic maps. The CS algorithm balances
rated into a population-based metaheuristic in LTH hybrid. between exploration and exploitation, while chaotic maps
HRH hybrid has sequential metaheuristics. Multiple self- update the control energy parameters to avoid a local
contained algorithms collaborate to find an optimal HTH optimum [34]. Another study called HBGWOHHO pro-
solution. This research uses the LTH paradigm to include posed to achieve a good balance while also improving the
the HHO exploration stage in the WOA algorithm. performance of the native GWO algorithm. The perfor-
Some hybrid algorithms have succeeded in improving mance of the search algorithm will be improved by prop-
the function selection than original algorithms [5]. The erly balancing exploration and exploitation [5]. For
very first time a hybrid metaheuristics method was used in tackling real-world optimization problems, an unique
feature selection was in the year 2000. The very first time a hybrid optimization method (H-HHONM) is suggested,
combination of metaheuristic search method was used in which combines the Nelder-Mead local search technique
selecting features was in the year 2000. [19], when a GA with the HHO [35]. The HMPA algorithm is a hybrid
algorithm with local search technique was developed to optimization method that divides the search agent popu-
control the search process. Another study on GA was lation into many sub-populations to balance exploration
undertaken to enhance the algorithm’s search by merging and exploitation capacities [36].
GA’s global search and the integrated regularization Despite major advances in attribute selection algo-
approach’s local search this method named hybrid genetic rithms, the No Free Lunch (NFL) theorem [37] in opti-
algorithm [20, 21] with wrapper-embedded feature mization argues that there is no such thing as an ideal
approach (HGAWE) [22]. In addition, several studies technique for solving all sorts of problems and never will
conducted to hybridize PSO algorithm for feature selection. be. Even with wrapper feature selection’s progress, better
For example, the hybrid PSO-SVM that intended to opti- solutions require more research and development. Fol-
mize the categorization task’s accuracy in low-dimensional lowing is a discussion of a hybrid wrapper-based method
dataset [23]. In contrast, the hybrid PSO method for high- for selecting features.
dimensional data combined the filter and wrapper approa-
ches to bring advantage of the best properties of both 2.1 Whale optimization algorithm (WOA)
methods, reducing computing time while increasing the
accuracy of the collected features proposed in [24]. In [25], WOA [12] is a Population-metaheuristic algorithm that
and PSO and GSA were combined to stabilize the algo- modeled the hunting behavior of humpback whales. The
rithm’s exploratory and exploitative characteristics. [26]. humpback whales pursue and feed on school of small fish
Similarly, WOA has been hybridized with other algo- or krill near the surface to survive. They create unique
rithms to improve the algorithm’s capability. For instance, maneuvers to paralyze the prey by swimming around them
the hybridization of the WOA with SA in order to obtain in a circle or a spiral shaped path. This method called
efficient search areas and improve the algorithm’s bubble-net attack. The hunting cycle of the humpback
exploitation stage [27]. In [28], a wrapper GWO and WOA whales is specified into two phases. The first phase is the
Feature Selection method proposed, and it aimed to reduce exploration stage when the whales are searching for the
the immature convergence and the local optima stagnation. prey randomly. The other phase is exploitation phase,
Furthermore, the HWOAFPA method [29] is created by where the whales are performing the bubble-net hunting.
combining the WOA with the flower pollination algorithm The WOA algorithm simulates the humpback whale
(FPA). This hybrid method begins by running the WOA searching strategy considering that the current position is
algorithm, and then the population of WOA is changed near to the best solution then the other agents update their
using the OBL during the process. Besides, a novel hybrid location with the finest search agent’s position. The fol-
filter-wrapper WOA model is proposed as a multi-objective lowing equations can be used to model this behavior
algorithm that optimizes both the filter and wrapper fitness mathematically:
functions at the same time [30]. A hybrid WOA modified  
! ! ! ! 
with Lévy flight and differential evolution (WOA-LFDE) D ¼  C : X ðtÞ  X ðtÞ ð1Þ
was developed in another work to solve JSSP [31]. ! ! !
Recently, a variety of hybridization technique applied X ðt þ 1Þ ¼ X ðtÞ  A :D ð2Þ
on HHO algorithm such as hybrid Sine Cosine Algorithm !
where X indicate the best obtained solution’s position, t
with Harris Hawks Optimization (SCHHO) [32], The main

123
Neural Computing and Applications

represents the number of iterations, the constant parameters serving as candidate solutions [37]. The system simulates
! ! Harris’ hawks’ perching strategies in two scenarios. Harris’
A , and C represented in Eqs. (3) and (4) as:
! hawks will perch in a variety of random locations around
A ¼ 2! a :!
r ! a ð3Þ their group home area in the first scenario. In Eq. (8), q 
! 0.5 represents this circumstance. The other possibility is
C ¼ 2:! r ð4Þ
that Harris’ hawks will reside nearby other members of the
where ! r are random vector between [0,1] and the family and prey. For q \ 0.5, Eq. (8) represents the
parameter !a is a random vector used to control the con- modelling of that option.

vergence process and it is decreasing linearly from 2 to 0 Xrand ðtÞ  r1 jXrand ðtÞ  2r2 XðtÞj; q  0:5
Xðt þ 1Þ ¼
throughout iterations. The value of !a can be calculated ðXrabbit ðtÞ  Xm ðtÞÞ  r3 ðLB þ r4 ðUB  LBÞÞ; q\0:5
using Eq. (5): ð8Þ
2 !
!
a ¼2t ð5Þ where X 1 ðt þ 1Þ denotes the next hawk location, X rand ðtÞ
T is a hawk chosen at random from the existing population,
where t is the current iteration and T is the maximum X ðtÞ indicates the contemporary vector hawk loca-
! tion,r 1 ,r 2 ,r 3 ,r 4 , and q represent records chosen at random
number of iterations the vector A is used to indicate the
within (0,1),), X rabbit ðtÞ denotes the prey’s location, X m is
transition process between the exploratory and exploitative
! the existing population of hawks’ average computed posi-
behaviors. When the absolute value of A is higher than 1 tion. Lastly, the search space’s lower and upper borders are
the search agents tend to keep exploring the search space or marked by the letters UB and LB, respectively.
to exploit the solution if the jAj\ 1. During the optimization process (Exploitation stage),
In WOA algorithm, the exploitation phase can be one of the hawk pursues the prey that has been spotted earlier in
two methods, shrinking encircling method or spiral-shaped the phase. The HHO algorithm has four options for simu-
method. In the shrinking encircling mechanism can be lating Harris’ hawks’ various attacking techniques.
achieved by decreasing the value of a in Eq. (2). The spiral Assuming that r denotes the likelihood of prey fleeing, an
path mechanism updates the distance between the most efficacious survive is denoted by r\0.5, and a failed effort
!
effective search agent obtained up to this point ( X ) and a to escape is denoted by r C 0.5. Depending on the likeli-
search agent (X) then update a new position for the search hood of the prey fleeing (r), to capture the prey, the hawks
agent according to the following Eq. (6): will get one of two methods: soft besiege or hard besiege.
0 The HHO algorithm parameter E is being used to identify
X ðt þ 1Þ ¼ D :ebl :cosð2plÞ þ X  ðtÞ ð6Þ the type of assaulting besiege. If the prey is unlikely to flee
To indicate which exploitative behavior will be per- (r C 0.5), moderate besiege will be employed when |E| C
formed, a probability value of 0.5 will be set for both. This 0.5, and severe besiege will be used if |E| \ 0.5. The
process modelled as: moderate besiege is mathematically represented by Eq. (9)
 0 bl and (10), whereas the severe besiege is mathematically
! D :e :cosð2plÞ þ X  ðtÞ; p  0:5 described by Eq (11):
X ð t þ 1Þ ¼ ~  ðt Þ  ~ ð7Þ
X A:D; p\0:5
Xðt þ 1Þ ¼ DXðtÞ  EjJXrabbit ðtÞ  XðtÞj ð9Þ
where p presents an arbitrary probability value between DðtÞ ¼ Xrabbit ðtÞ  XðtÞ ð10Þ
[24].
Xðt þ 1Þ ¼ Xrabbit ðtÞ  EjDXðtÞj ð11Þ
2.2 Harris Hawks Optimization (HHO) where X ðt þ 1Þ denotes the difference between the prey’s
position and the most recent location, DX ðtÞ represents an
HHO is a population-based metaheuristic algorithm [9] arbitrary number within (0,1), and JX rabbit ðtÞ denotes the
inspired on Harris Hawks’ ‘‘Seven Kills’’ hunting strategy. escaping prey’s irregular bounce tendency during the
The task of hunting could take only a few seconds or evasion operation. Finally, the J value can vary at irregular
several hours, relying on the prey’s characteristics evasion intervals in each iterative process to replicate the charac-
techniques. Just like any other metaheuristics, the search teristic of prey movements.
procedure is divided into two sections in his algorithm: When the prey has a probability of effectively evasion r
exploration and exploitation. The process of perching and \ 0.5, light encircle with gradual speedy leap arises when
exploring is referred to as the exploration stage. The |E| C 0.5 Eqs. (12), (14), (15), and (16), while the rough
desired or ideal prey serves as the best candidate solution in encircle with gradual speedy leap occurs when |E| \ 0.5
each phase of the HHO algorithm, with Harris’ hawks Eqs. (13), (14), (15), and (16).

123
Neural Computing and Applications

Y ¼ X rabbit ðtÞ  EjJX rabbit ðtÞ  X ðtÞ ð12Þ the exploitation phase is based on BWOA employing
Eq. (2) or (6), depending on the humpback whale’s
Y ¼ X rabbit ðtÞ  EjJX rabbit ðtÞ  X m ðtÞj ð13Þ
exploitative manner. The created algorithm (BWOAHHO)
Z ¼ Y þ S  LFðDÞ ð14Þ has a better chance of avoiding local optimal solutions,
 which improves solution accuracy. The proposed
Y; if f ðYÞ\FðXðtÞÞ
X ð t þ 1Þ ¼ ð15Þ hybridization method, shown in Fig. 1, was applied to
Z; if f ðZÞ\FðXðtÞÞ
improve the BWOA search strategy. The variant model
0   1b1
pb (BWAOHHO) starts by setting the value of the coefficient
ur Cð 1 þ b Þ  sin 2
LF ð xÞ ¼ 0:01  @   A vector A.
1 ;r ¼ b1
jvjb C 2 b2 2 Þ
1þb ð If |p|C 0.5, the whales will use the shrinking encircling
approach to update their whereabouts, as modeled in
ð16Þ
Eq. (2). The exploration phase will begin when |p|\ 0.5
where D indicates the problem’s size, S is an arbitrary and |A|[ 1 are equal, and the search agents will use one of
parameter denoted by the size 19D, LF represents levy two tactics. If we assume that each strategy has an equal
flight component, and it is deduced from Eq. (16), and Y chance r. For the situation of r \ 5 in Eq. (19), as well as
are Z utilized from Eqs. (12), (13). Whereas U and v are the other Strategic Condition of r  0.5the search agent is
randomly initialized parameters within (0,1), and b denotes placed near the other group members, the position is ran-
a standard parameter given a value of 1.5. domly chosen within the members range defined in
Eq. (19).

! X r ðtÞ  C jX r ðtÞ  2r 1 X ðtÞj; r  0:5
X ð t þ 1Þ ¼
3 The hybrid algorithm (BWOAHHO) ðX P ðtÞ  X m ðtÞÞ  CðLB þ r 2 ðUB  LBÞÞ; r\0:5
ð19Þ
The search space for feature selection is binary or, in other !
words, has two boundaries: zero and one. The WOA C ¼ 2:! r3 ð20Þ
algorithm’s original search space is continuous. As a result, !
where X ðt þ 1Þ. is the next stage for locating the search
the feature selection problem cannot be solved using native
agents, X r ðtÞ represent the current population’s random
WOA. An improved variety of the optimization technique
position for search agents, X P ðtÞ specifies the prey’s loca-
must be created (binary version). To update search agent
tion, X m ðtÞ reflects the search agents’ average position, C
placements for the WOA algorithm on a binary search
denotes a stochastic factor derived from Eq. (20), r, r1, r2
region, the subsequent sigmoidal equation has been
are arbitrary coefficients 2[0–1]. To prevent local optima,
included:
the value C shows a high level of unpredictability in the
1 algorithm. The parameters r, r1 and r2 are scaling coeffi-
Sð x Þ ¼ ð17Þ
1 þ eð10ðx0:5ÞÞ cients that increase the randomness with which the method
 generates random positions within feature selection
1; sigmoid X   r
Xbinary ðt þ 1Þ ¼ f ð xÞ ¼ ð18Þ boundaries.
0; otherwise
Once |A|\ 1 is true the foraging agents must update their
where X binary ðt þ 1Þ, is a binary location update technique positions according to the spiral-shaped method using
sigmoidal equation SðxÞ characterized in Eq. (17). The !
Eq. (6). Based on the absolute value of the operator A ,
parameter r indicates a randomly selected number drawn !
from a consistent distribution 2 ½0; 1 since the value of A is a random parameter ranging
BWOA is a relatively new optimization technique which from - 2a to 2a, the search agents will switch between the
produces competitive results in comparison to other explorative and the exploitative behavior until the maxi-
metaheuristic algorithms like PSO, FEP, GSA and DE [12]. mum number iterations is exceeded. Once jAj\1 that sit-
However, the native BWOA algorithm function an operator uation indicates that the superlative search applicant ‘‘best
to control the search in the exploration phase [37]. In the solution’’ is found (exploitation), while if |Aj  1 that
current study, the BWOA exploration phase will be indicates an arbitrary search applicant is selected to
replaced by an improved one. This is a hybridization investigate a new area. The |A| is intended by Eq. (21).
! !
between algorithms BWOA and HHO. In this case of jA j ¼ j2!a :!r 1  aj ð21Þ
hybridization, the exploration phase from HHO is embed-
ded in the algorithm WOA to increase the randomness of where ! a is a value randomly selected and, over the iter-
the optimum solution search. As a result, the given algo- ations, it drops linearly from 2 to 0.
rithm’s exploration phase is based on HHO Eq. (12), while

123
Neural Computing and Applications

Fig. 1 Scheme of the proposed


process of hybridization

123
Neural Computing and Applications

As in Algorithm 1, the pseudo-code of the hybrid where ER denotes the margin of error, Sf is the number of
BWOAHHO is obtained. features specified, and T f is the entire number of attributes,

The purpose of selected attributes is to limit the per- and = [0, 1] is a constant to manage accuracy [42].
centage of characteristics and error rates of classification
[35]. In other words, by removing unneeded and duplicated 3.1 BWOAHHO potential benefits
traits and preservation of only those of use the accuracy of
the classifications is improved [33]. The KNN classifier has The purpose of this research is to strengthen the exploring
been utilized because fitness function is easy to evaluate stage of WOA in order to improve Classification quality
[38, 39, 40]and [41]. The fitness function used has been and execution time reduction. It is expected that combining
conveyed in Eq (22). these two heuristic methods will result in a simplified,
functional hybrid algorithm.
jSf j
fitness ¼ aðERÞ þ ð1  aÞ ð22Þ
jTf j

123
Neural Computing and Applications

3.1.1 Computational complexity The value of K is determined through observation and


experimentation. The value of k = 5 was chosen for the
The computational complexity of the BWOAHHO is built experiments, since it performed best across all datasets
around three key components: initialization, fitness evalu- [38, 44]. The K-fold cross-validation method was used to
ation, and the population size updating process. Because partition the datasets into K segments [45]. One segment
there are N search operators, the complexity of processing assigned to be a test segment, whereas the rest (K-1) serve
in the initial stage is O (N). It entails determining the fittest as training segments [46].
solution and revising the positions among all search engi- Until the best solution is found, the segments of train
nes as part of the update process. As a result, the updating and test datasets have been rearranged and replicated at
procedure is O (T x N) ? O (T x N x B), where B indicates random [45]. According to [47], the number of iterations
the limit of search and T denotes the total amount of (T) and the size of the population (N) are two important
iterations. The new solution’s computing challenge criteria that determine the applicability of a given method.
(BWOAHHO) is O (N x (T ? TD ? 1). The variables k, N, UB, LB, and T have been speci-
fied according to the conventional parameter settings
specified in [37] and [44]. The parameters for the proposed
4 Results of experiment BWOAHHO algorithm are listed in Table 2. An Intel (R)
CoreTMi7-6700 machine with 3.4GHz CPU and 8GB
4.1 Used datasets RAM was used to run the method. The results of the
original codes BWOA from [42] and BHHO from [43]
The benchmark datasets used to implement the research were compared after they were both run on the same sys-
work BWOAHHO’s practicability are listed in Table 1. tem. The BPSO [48], BGA [4], and WOASAT-2 algo-
These datasets were retrieved through UCI’s machine rithms [33] were also compared to the BWOAHHO
learning library [43]. algorithm. Accuracy of classification, average fitness,
optimal fitness, least fitness, the multitude of attributes that
4.2 Measurement configurations have been selected, and the time required for computation
were used to analyze and compare the results.
The best solutions were assessed using the K-nearest
neighbor (KNN) method with Euclidean distance and k = 5. 4.3 Metrics of evaluation

To assess the efficiency of the created solutions, we used


Table 1 Datasets for benchmark testing
data analysis indicators such as mean accuracy, average
# Dataset Attributes number Instances number fitness, optimal fitness, least fitness, Average attribute size
1 Breastcancer 9 699
selected, and mean execution time.
2 BreastEW 30 569
4.3.1 Mean accuracy
3 CongressEW 16 434
4 Exactly 13 1000
The statistic measures the classification algorithm’s accu-
5 Exactly2 13 1000
racy in identifying the most valuable attributes. Equa-
6 HeartEW 13 270
tion (23) is used to calculate the average accuracy of the
7 IonosphereEW 34 351
algorithm after it has been performed M times:
8 KrvskpEW 36 3196
9 Lymphography 18 148
10 M-of-N 13 1000 Table 2 BWOAHHO measurement configuration
11 PenglungEW 325 73
Component Value
12 SonarEW 60 208
13 SpectEW 22 267 Size of the population (N) 10
14 Tic-tac-toe 9 958 Upper selection bound (UB) 1
15 Vote 16 300 Lower selection bound (LB) 0
16 WaveformEW 40 5000 The total iterations number (T) 100
17 WineEW 13 178 Multitude of recurring runs (M) 20
18 Zoo 16 101 Fitness factor value (a) 0.99

123
Neural Computing and Applications

Table 3 The proposed


# Dataset Accuracy Fitness Selected attributes Execution time
BWOAHHO yielded the
following results. Average (Av) Av. Std. Av. Std. Av. Std. Av. Std.
results and the standard
deviation (Std) algorithm are 1 Breastcancer 0.99 0.01 0.01 0.01 4.6 1.33 6.22 0.92
used to represent the data 2 BreastEW 0.99 0.01 0.03 0.05 18.1 3.78 5.18 0.72
3 CongressEW 0.98 0.01 0.02 0.01 5.3 2.68 5.04 1.45
4 Exactly 0.85 0.07 0.15 0.07 8.7 1.62 9.29 2.45
5 Exactly2 0.76 0.01 0.23 0.01 5 2.29 6.18 1.09
6 HeartEW 0.89 0.03 0.15 0.18 7.2 2.35 5.41 0.29
7 IonosphereEW 0.93 0.02 0.07 0.01 7.7 3.04 5.11 0.89
8 KrvskpEW 0.97 0.01 0.04 0.01 30.2 5.53 17.92 1.92
9 Lymphography 0.89 0.03 0.11 0.02 9.2 2.97 4.85 0.36
10 M-of-n 0.94 0.02 0.09 0.13 9.9 1.24 7.92 1.17
11 PenglungEW 0.96 0.03 0.04 0.03 123.3 43.64 6.73 1.72
12 SonarEW 0.93 0.03 0.07 0.03 25 10.83 8.68 3.34
13 SpectEW 0.91 0.02 0.10 0.02 8.7 1.88 5.01 0.24
14 Tic-tac-toe 0.85 0.01 0.14 0.03 6.8 0.87 6.61 0.46
15 Vote 0.97 0.01 0.02 0.01 5.1 3.78 5.11 0.60
16 WaveformEW 0.82 0.01 0.16 0.01 32.6 3.41 17.26 2.48
17 WineEW 1.00 0.01 0.00 0.01 7.7 2.10 5.03 0.35
18 Zoo 0.98 0.02 0.02 0.02 10.2 1.33 4.86 0.33

Table 4 Average classification


Dataset Proposed BWOAHHO BWOA BHHO BGA BPSO WOASAT-2
accuracy of proposed
BWOAHHO and relevant work Breastcancer 0.99 0.98 0.98 0.96 0.95 0.97
approaches
BreastEW 0.99 0.97 0.98 0.94 0.94 0.98
CongressEW 0.98 0.96 0.98 0.94 0.94 0.98
Exactly 0.85 0.78 1.00 0.67 0.68 1.00
Exactly2 0.76 0.77 0.79 0.76 0.75 0.75
HeartEW 0.89 0.88 0.91 0.82 0.78 0.85
IonoshpereEW 0.93 0.92 0.91 0.83 0.84 0.96
KrvskpEW 0.97 0.97 0.98 0.92 0.94 0.98
Lymphography 0.89 0.88 0.88 0.71 0.69 0.89
M-of-n 0.94 0.95 1.00 0.93 0.86 1.00
PenglungEW 0.96 0.89 0.94 0.70 0.72 0.94
SonarEW 0.93 0.91 0.96 0.73 0.74 0.97
SpectEW 0.91 0.87 0.90 0.78 0.77 0.88
Tic-tac-toe 0.85 0.84 0.83 0.71 0.73 0.79
Vote 0.97 0.96 0.97 0.89 0.89 0.97
WaveformEW 0.83 0.83 0.84 0.77 0.76 0.76
WineEW 1.00 0.99 1.00 0.93 0.95 0.99
Zoo 0.98 0.96 0.97 0.88 0.83 0.97
Average 0.92 0.90 0.93 0.83 0.82 0.92

1 XM 4.3.2 Average fitness


Average Accuracy ¼ AccuracyK ð23Þ
M k¼1
The average value of the attained fitness function is
where AccuracyK denotes the accuracy at run k.
referred to by this metric. Equation (24) is used to calculate
the algorithm’s mean fitness after it has been performed M
times:

123
Neural Computing and Applications

Table 5 Comparison of the


Dataset Proposed BWOAHHO BWOA BHHO BGA BPSO WOASAT-2
proposed BWOAHHO’s mean
fitness function and a similar Breastcancer 0.01 0.03 0.02 0.03 0.03 0.04
works technique
BreastEW 0.03 0.04 0.01 0.04 0.03 0.03
CongressEW 0.02 0.06 0.02 0.04 0.04 0.03
Exactly 0.15 0.23 0.05 0.28 0.28 0.01
Exactly2 0.23 0.23 0.22 0.25 0.25 0.25
HeartEW 0.15 0.12 0.09 0.14 0.15 0.16
IonosphereEW 0.07 0.08 0.08 0.13 0.14 0.04
KrvskpEW 0.04 0.03 0.03 0.07 0.05 0.02
Lymphography 0.11 0.12 0.11 0.17 0.19 0.11
M-of-n 0.09 0.07 0.01 0.08 0.11 0.01
PenglungEW 0.04 0.11 0.04 0.22 0.22 0.06
SonarEW 0.07 0.14 0.05 0.13 0.13 0.03
SpectEW 0.10 0.13 0.08 0.14 0.13 0.13
Tic-tac-toe 0.14 0.15 0.15 0.24 0.24 0.21
Vote 0.02 0.09 0.02 0.05 0.05 0.04
WaveformEW 0.16 0.17 0.16 0.20 0.22 0.25
WineEW 0.00 0.01 0.00 0.01 0.02 0.01
Zoo 0.02 0.04 0.03 0.08 0.10 0.04
Average 0.08 0.10 0.07 0.13 0.13 0.08

1 XM k statistic. The average selection is obtained as in Eq. (27)


Mean fitness ¼ g
k¼1 
ð24Þ
M after the procedure has been executed M times:
where gk indicates set to be the fitness function value over 1 XM Avg:sizek
Average selection ¼ ð27Þ
the operated k. M k¼1 Tf
where T f represents the set of attributes, whereas Avg.sizek
4.3.3 Optimal fitness
denotes the mean size of the attribute values over the
operated k.
This measure indicates the fitness function’s minimum
value. After running the algorithm M times, the best fitness
4.3.6 Mean execution time
is calculated as shown in Eq. (25):
Best fitness ¼ MinM k
K¼1 g ð25Þ The average computing time is expressed in seconds. The
gk
where denotes the best fitness solution obtained over the average computing time is computed using Eq. (28)
XM
operated k. Average time ¼ Avg:timek ð28Þ
k¼1

4.3.4 Least fitness where Avg:timek is the average amount of time spent
computing over the operated k.
The mentioned measure relates to the fitness function’s
greatest value. The worst fitness of the algorithm is cal-
culated as in Eq. (26) when it is executed M times: 5 Discussion of the findings
Worst fitness ¼ MaxM k
K¼1 g ð26Þ
The proposed hybrid algorithm’s outcomes are discussed in
where gk denotes the best fitness value over the operated k. this section.

4.3.5 Average attribute size selected 5.1 BWOAHHO algorithm

The mentioned ratio of the average size of the selected The results of utilizing the new BWOAHHO heuristic
features to the total number of features is referred to as this method to 18 datasets are stated in Table 3. The findings
include average variable accuracy, mean fitness, selection

123
Neural Computing and Applications

Table 6 Comparison of the


Dataset Proposed BWOAHHO BWOA BHHO BGA BPSO WOASAT-2
introduced BWOAHHO and
relevant work methods optimal Breastcancer 0.01 0.01 0.00 0.02 0.03 0.03
fitness function
BreastEW 0.01 0.011 0.00 0.02 0.02 0.02
CongressEW 0.01 0.07 0.01 0.03 0.03 0.02
Exactly 0.20 0.13 0.001 0.27 0.21 0.01
Exactly2 0.20 0.23 0.21 0.22 0.22 0.23
HeartEW 0.07 0.14 0.06 0.12 0.13 0.13
IonosphereEW 0.09 0.07 0.06 0.09 0.12 0.03
KrvskpEW 0.03 0.03 0.02 0.03 0.03 0.02
Lymphography 0.07 0.07 0.09 0.12 0.14 0.09
M-of-n 0.07 0.06 0.00 0.02 0.06 0.01
PenglungEW 0.07 0.07 0.00 0.13 0.13 0.03
SonarEW 0.02 0.07 0.02 0.07 0.07 0.01
SpectEW 0.06 0.13 0.04 0.12 0.10 0.11
Tic-tac-toe 0.13 0.15 0.11 0.21 0.21 0.20
Vote 0.01 0.02 0.00 0.03 0.03 0.02
WaveformEW 0.17 0.17 0.15 0.19 0.21 0.23
WineEW 0.00 0.00 0.01 0.00 0.00 0.00
Zoo 0.00 0.01 0.00 0.00 0.03 0.00
Average 0.07 0.08 0.04 0.09 0.10 0.07

Table 7 Comparison of the


Dataset Proposed BWOAHHO BWOA BHHO BGA BPSO WOASAT-2
proposed BWOAHHO and
comparable work techniques Breastcancer 0.01 0.036 0.04 0.03 0.03 0.04
using the worst fitness function
BreastEW 0.02 0.04 0.02 0.04 0.03 0.04
CongressEW 0.03 0.07 0.05 0.04 0.04 0.05
Exactly 0.20 0.26 0.12 0.28 0.28 0.01
Exactly2 0.20 0.24 0.24 0.25 0.25 0.27
HeartEW 0.09 0.16 0.14 0.14 0.15 0.18
IonosphereEW 0.1 0.09 0.13 0.13 0.14 0.05
KrvskpEW 0.05 0.04 0.05 0.07 0.05 0.02
Lymphography 0.10 0.11 0.14 0.17 0.19 0.14
M-of-n 0.09 0.07 0.02 0.08 0.11 0.01
PenglungEW 0.07 0.21 0.18 0.22 0.22 0.11
SonarEW 0.02 0.15 0.12 0.13 0.13 0.05
SpectEW 0.15 0.15 0.13 0.14 0.13 0.15
Tic-tac-toe 0.15 0.17 0.21 0.24 0.24 0.23
Vote 0.02 0.05 0.06 0.05 0.05 0.04
WaveformEW 0.20 0.17 0.18 0.20 0.22 0.26
WineEW 0.02 0.03 0.02 0.01 0.01 0.03
Zoo 0.05 0.05 0.09 0.03 0.10 0.10
Average 0.09 0.12 0.11 0.13 0.13 0.10

size of the feature and computing time. The method ran percent accurate. The Breastcancer dataset was the smallest
twenty times for each dataset. WineEW was the dataset in terms of feature selection (4.6 features). Following that,
with the greatest reported average accuracy of 100%. the Exactly2 and Vote datasets, which had feature sizes of
Following that, in the BreastEW and Breastcancer datasets, 5 and 5.1, respectively. It is worthwhile to mention that the
the average accuracy was 99 percent. The CongressEW and PenglungEW, Zoo, SpectEW, and WineEW datasets took
Zoo datasets both had a 98 percent accuracy rate. On the least amount of time to compute, with times of 4.8,
average, the two datasets, Vote and KrvskpEW, were 97 4.86, 5.01, and 5.04 seconds, respectively.

123
Neural Computing and Applications

Table 8 Comparison of the


Dataset Proposed BWOAHHO BWOA BHHO BGA BPSO WOASAT-2
proposed BWOAHHO’s
average selected attributes size Breastcancer 4.60 6.1 6.8 5.1 5.7 4.2
and related work techniques
BreastEW 18.10 18.9 21.2 16.3 16.6 11.6
CongressEW 5.30 6.2 9.2 6.6 6.8 6.4
Exactly 8.7 10.2 6.8 10.8 9.8 6.0
Exactly2 5 6.35 8 6.2 6.2 2.8
HeartEW 7.2 8.8 8.8 9.5 7.9 5.4
IonosphereEW 7.7 6.55 18.1 17.3 19.2 12.8
KrvskpEW 30.2 30.5 35.85 22.4 20.8 18.4
Lymphography 9.2 9 10.6 11.0 9.0 7.2
M-of-n 9.9 9.6 6.9 6.8 9.1 6.0
PenglungEW 123.3 129.9 166.5 177.1 178.8 127.4
SonarEW 25 36.1 37.4 33.3 31.2 26.4
SpectEW 8.7 11.9 12.8 11.7 12.5 9.4
Tic-tac-toe 6.8 8.8 9.0 6.8 6.6 6.0
Vote 5.1 8.7 9.1 6.6 8.8 5.2
WaveformEW 32.6 35.5 27.9 25.30 22.7 20.6
WineEW 7.7 8.3 7.8 8.6 8.4 6.40
Zoo 10.2 10.7 10 21.8 21.66 5.60
Average 18.07 20.12 22.3 22.40 22.32 15.99

5.2 Comparison of the proposed BWOAHHO (execution) time should be considered while evaluating the
algorithm to other algorithms algorithm’s performance.

The BWOAHHO method is proposed to provide extra 5.2.2 Mean fitness comparison
unpredictability to the BWOA algorithm’s exploration
phase search process. In terms of Accuracy of classifier, Table 5 reveals that the novel BWOAHHO has an average
average fitness, optimal fitness, least fitness, Attribute size fitness of 0.08, which is lower than the average fitness of
selected, and the time it takes to compute, Tables 4 through BWOA (0.10), BPSO (0.13), and BGA (0.13). However,
9 compare the proposed BWOAHHO to several previous the proposed BWOAHHO algorithm showed best average
wrapper-based methods. fitness results in ten over eighteen datasets. Low fitness
levels suggest that an algorithm is performing well. In 14
5.2.1 Classification accuracy comparison of the 18 datasets studied, the BWOAHHO algorithms
outperformed the BWOA algorithms in terms of fitness.
In Table 4, the developed BWOAHHO algorithm is com- Furthermore, the suggested BWOAHHO method outper-
pared to related metaheuristic algorithms with regards to formed the BPSO and BGA algorithms on all 18 datasets
accuracy. BWOAHHO had a classification accuracy of examined. It also outperformed WOASAT-2 in twelve
0.92 in the 18 datasets, which was greater than BWOA datasets tested. It’s worth noting that BHHO’s average
(0.90), BPSO (0.82), and BGA (0.82). (0.83). The proposed classification accuracy was 0.06, which was somewhat
approach is a hybridization of BWOA and BHHO, two higher than the proposed hybrid algorithm’s average
binary algorithms. Out of the 18 datasets analyzed, the fitness.
BWOAHHO algorithms outscored the BWOA algorithms
in 14 of them, demonstrating that hybridization improved 5.2.3 Best fitness comparison
classification accuracy. Furthermore, the proposed
BWOAHHO method outperformed the BPSO and BGA As mentioned in Table 6, the average best fitness of the
algorithms on all 18 datasets examined. It also outper- proposed BWOAHHO, BWOA, BHHO, BPSO, BGA and
formed WOASAT-2 in 13 of the 18 datasets tested. It’s WOSAT-2 are 0.07, 0.08, 0.04, 0.09 and 0.08 respectively.
worth noting that BHHO’s average classification accuracy The experiments result of best fitness indicates that the
was 0.93, which was somewhat higher than the proposed proposed algorithm outperformed the BWOA algorithm in
algorithm’s average accuracy. However, other variables fifteen datasets, whereas outperformed BPSO and BGA
like as fitness, selected feature size, and computing algorithms in seventeen datasets. However, the average of

123
Neural Computing and Applications

Fig. 2 BWOAHHO and other 1


algorithms achieved a selection
0.9
accuracy and b the quantity of
selected attributes in limited 0.8
datasets 0.7
0.6

Accuracy
0.5
0.4
0.3
0.2
0.1
0
Breastcancer Exactly Exactly 2 HeartEW M-of-n Tic-tac-toe WineEW
Datasets

BWOAHHO BWOA BHHO BPSO BGA WOASAT-2


(a)

WineEW

Tic-tac-toe

M-of-n
Dataset

HeartEW

Exactly2

Exactly

Breastcancer

0 3 6 9 12 15
Feature Size

WOASAT-2 BGA BPSO BHHO BWOA BWOAHHO


(b)

the optimal fitness of native BHHO algorithm is slightly 5.2.5 Selected features comparison
higher than the proposed BWOAHHO.
Another statistic to consider when evaluating a feature
5.2.4 Worst fitness comparison selection algorithm is its capacity to reduce the number of
selected characteristics in order to eliminate redundant
Furthermore, based on the average worst fitness value, ones. In 18 datasets, Table 8 compares the selected feature
Table 7 illustrated that the new metaheuristic BWOAHHO size achieved by the proposed BWOAHHO algorithm with
was not the worst in all datasets. The average of the worst that acquired by the other five techniques. BWOAHHO,
fitness of proposed BWOAHHO is 0.09, which is the best BWOA, BHHO, BGA, and WOASAT-2 achieved average
value over BWOA (0.12), BHHO (0.11), BGA (0.13), selected feature sizes of 18.07, 20.12, 22.3 22.32, 22.40,
BPSO (0.13) and WOASAT-2 (0.10). and 15.99, respectively. In other words, BWOAHHO out-
performed all algorithms except WOASAT-2 in terms of
specified feature size.

123
Neural Computing and Applications

Fig. 3 BWOAHHO and other 1


algorithms achieve a selection
0.9
accuracy and b the quantity of
selected attributes 0.8
in intermediate datasets
0.7

0.6

0.5

Accuracy
0.4

0.3

0.2

0.1

0
CongressEW Lymphography SpectEW Vote Zoo
Dataset
BWOAHHO BWOA BHHO BPSO BGA WOASAT-2

(a)

Zoo

Vote
Dataset

SpectEW

Lymphography

CongressEW

0 5 10 15 20 25
Feature Size

WOASAT-2 BGA BPSO BHHO BWOA BWOAHHO

(b)

5.2.6 Computational time comparison takes considerably less time to compute (in seconds) than
native BWOA and other algorithms.
Table 8 shows how long the proposed BWOAHHO
method, and the other three algorithms take to compute in 5.3 Performance of algorithms at various data
seconds. For BWOAHHO, BWOA, BHHO, and WOA- sizes is compared.
SAT-2, the average computational time was 7.36, 7.81,
27.45, and 167.44 s, respectively. It is clear that the pro- On datasets of various sizes, the proposed technique was
posed BWOAHHO method reduces the processing time further evaluated in contrast to existing algorithms. The 18
significantly. The proposed method has clearly offered a datasets investigated were classified into three categories,
robust feature selection procedure, as evidenced by the following [54, there are three types of datasets: limited
statistical findings. The proposed BWOAHHO demon- datasets (0–15 attributes), intermediate datasets (16–25
strated a remarkable capacity to choose fewer character- attributes), and big or enormous volumes of data (26?
istics while keeping excellent accuracy. Furthermore, attributes) (30–350 attributes). In each of the three cate-
across all 18 datasets examined, the proposed BWOAHHO gories under consideration, it was assessed whether the
suggested BWOAHHO algorithm and other algorithms

123
Neural Computing and Applications

Fig. 4 BWOAHHO and other 1


algorithms achieve a selection
accuracy and b the quantity of 0.9
selected attributes in big
0.8
datasets
0.7

0.6

Accuracy
0.5

0.4

0.3

0.2

0.1

0
BreastEW IonosphereEW KrvskpEW PenglungEW SonarEW WaveformEW
Dataset

BWOAHHO BWOA BHHO BPSO BGA WOASAT-2

(a)

WaveformEW

SonarEW

PenglungEW
Dataset

KrvskpEW

IonosphereEW

BreastEW

0 20 40 60 80 100 120 140 160 180 200


Feature Size

WOASAT-2 BGA BPSO BHHO BWOA BWOAHHO

(b)

increased the accuracy of classification and the size of the outperformed BPSO by a small margin. In addition, in five
specified attributes. datasets, the proposed algorithm exceeded WOASAT-2
Figure 2 compares the classification accuracy and fea- when it came to classification accuracy. In two datasets,
ture selected size of the proposed BWOAHHO algorithm (Breastcance and WineEw, BWOAHHO outscored the
with the accuracy and feature selected size of the other five BHHO metaheuristic when it comes to accuracy.
algorithms in the category of small datasets. As demon- BWOAHHO selected a smaller subset of attributes across
strated in Fig. 2a, BWOAHHO outperformed native all datasets when contrasted to BWOA and BGA, with the
BWOA in five datasets (breast cancer, Exactly, HeartEW, exception of the M-of-n dataset, as shown in
Tic-tac-toe, and WineEW) out of seven datasets. It also Fig. 2(b) BWOAHHO selected lesser number of features
outperformed BPSO and BGA in terms of accuracy in all than BHHO in four datasets, Tic-tac-toe, HeartEW,
datasets with the exception of Exactly2, where BGA Extactly2 and Breastcancer. When compared to BPSO, the

123
Neural Computing and Applications

Table 9 Comparison of
Dataset Proposed BWOAHHO BWOA BHHO WOASAT-2
proposed BWOAHHO with
BWOA, BHHO, and Breastcancer 6.22 11.60 29.97 43.24
WOASAT-2 execution times (in
seconds) BreastEW 5.18 7.32 27.07 43.18
CongressEW 5.04 5.18 16.63 37.34
Exactly 9.29 5.98 18.80 57.39
Exactly2 6.18 6.55 31.97 55.85
HeartEW 5.41 5.73 24.49 28.67
IonosphereEW 5.11 5.41 25.20 30.79
KrvskpEW 17.92 22.75 35.85 641.01
Lymphography 4.85 5.44 19.24 27.15
M-of-n 7.92 6.24 31.90 52.32
PenglungEW 6.73 5.43 27.46 28.47
SonarEW 8.68 5.05 25.29 28.02
SpectEW 5.01 5.68 20.72 30.34
Tic-tac-toe 6.61 6.90 23.92 51.93
Vote 5.11 5.55 20.24 30.63
WaveformEW 17.26 19.08 75.21 1770.48
WineEW 5.03 5.46 21.33 29.59
Zoo 4.86 5.18 22.49 27.55
Average 7.36 7.81 27.45 167.44

Table 10 ANOVA results summary BPSO, BGA, and BHHO in all of medium-sized datasets.
Groups Count Sum Average accuracy Variance
Figure 3 shows that BWOAHHO selected a smaller subset
of features when compared to all other algorithms across
BWOAHHO 18 16.62 0.92 0.004459 all medium-sized datasets, with the exception of
BWOA 18 16.31 0.90 0.004649 Lymphography.
BHHO 18 16.84 0.93 0.004238 Figure 4 contrasts the novel BWOAHHO method to
BPSO 18 14.76 0.82 0.009271 other five algorithms in the large dataset’s category. In all
BGA 18 14.87 0.82 0.009849 large datasets except KrvskpEW, BWOAHHO outper-
WOASAT-2 18 16.63 0.92 0.007002 formed native BWOA, as shown in Fig. 2a. It also sur-
passed BPSO and BGA on large datasets in the matter of
accuracy. The new method also outperformed WOASAT-2
with regard to accuracies in two datasets (BreastEW and
Table 11 ANOVA results in Detail WaveformEW). In five out of six datasets, BWOAHHO
produced almost identical accuracy as BHHO. Notably, the
Source of variation SS df MS F P-value
suggested BWOAHHO achieved the best accuracy in the
Between groups 0.244 5 0.049 7.44 5.62E-06 largest dataset (PenglungEW, with 325 features). Figure 4b
Within groups 0.670 102 0.007 shows that BWOAHHO selected a smaller subset of fea-
Total 0.915 107 tures than other methods in two large datasets (SonarEW
and PenglungEW). Among the other large datasets,
WOSAT-2 had the fewest selected features. However, in
the largest dataset (PenglungEW), our proposed
new algorithm selected a lesser number of features in four BWOAHHO reduced the number of features the best.
datasets (WineEW, HeartEW, Exactly2 and Breastcancer) Across a variety of dataset sizes, the newly developed
out of the seven small size datasets. BWOAHHO algorithm achieves good results and, in many
Figure 3 compares the classification accuracy and fea- cases, outperforms current state-of-the-art methods.
ture selection size of the proposed BWOAHHO algorithm A hybrid approach, as known in which at least two
with that of other five algorithms in the medium dataset procedures are merged to improve the performance of each
category. Figure 3a shows that the classification accuracy technique, is one option for achieving this balance.
of BWOAHHO was higher than that of native BWOA, Hybridizing the WOA algorithm with the HHO algorithm
improves the balance between global search and local

123
Neural Computing and Applications

search, which leads to an enhancement in the accuracy of BWOAHHO. It’s fair to say that the proposed hybrid
the new hybrid algorithm. Besides, the hybrid BWOAHHO method’s findings improve the native BWOA’s perfor-
algorithm showed a significant improvement in the exe- mance. In terms of accuracy, selected attributes size, and
cution time. execution time, the novel hybrid technique outperformed
the native BWOA algorithm. These findings show that the
5.4 Significance analysis new hybridization strategy significantly enhanced the
native BWOA algorithm’s performance. Furthermore,
A one-way ANOVA measure was applied to examine the when compared to BHHO, BPSO, BWOAHHO produced a
significance of the disparities in terms of statistics in the significant improvement, as specified by the p-value of
produced accuracy attained through the described fewer than 5%.
BWOAHHO vs. further algorithms. The goal of this Future research project could be to use a hybrid WOA
inquiry is to determine whether the results from the two algorithm to solve real issues like health data. Another
assessments are different. We compared BWOAHHO’s future projects, we could combine the WOA algorithm with
results to those of state-of-the-art approaches on eighteen another P-metaheuristic algorithm, such as GSA.
datasets. The discussion has been framed in terms of two
Acknowledgements Research reported in this publication was sup-
opposing statistical hypotheses by theoretical and applied
ported by Fundamental Research Grant Project (FRGS) from the
ecologists: we denote that there is no difference between Ministry of Education Malaysia (FRGS/1/2018/ICT02/UTP/03/1)
the model’s population parameters (i.e., the null hypothe- under UTP Grant No. 015MA0-013.
sis, H0), which assumes that there is no significant differ-
Data availability The data were taken from UCI public repository.
ence in the accuracy of the proposed BWOAHHO
optimizer and other optimizers, whereas the other indicates Code availability Will be made available upon request.
either one deterministic or probabilistic alternative (i.e., the
alternative hypothesis, Ha) that proves the contrary Declarations
hypothesis that there is a substantial difference in the
accuracy of the proposed BWOAHHO and the other opti- Conflict of interest The author declare that they have no conflict of
mizers. Significant levels of implication more than 5% interest.
confirm the null hypothesis (h0) statement, and conversely.
The methods that were employed throughout this evalua-
tion are similar to those used in [49, 50] and [51]. The
References
ANOVA results are summarized in Tables 9 and 10. The 1. Xue B, Zhang M, Browne WN, Yao X (2016) A survey on
p-values shown indicate that. It can be shown that the evolutionary computation approaches to feature selection. IEEE
suggested BWOAHHO outperformed the BWOA signifi- Trans Evol Comput 20(4):606–626. https://doi.org/10.1109/
cantly, with a p-value of 5.62E-06. Furthermore, when TEVC.2015.2504420
2. Gou J, Xue Y, Ma H, Liu Y, Zhan Y, Ke J (2020) Double graphs-
compared to BHHO, BPSO, BGA, and WOASAT-2, based discriminant projections for dimensionality reduction. Neur
BWOAHHO provided a significant improvement. The total Comput Appl. https://doi.org/10.1007/s00521-020-04924-5
number of data points (N) used is 18. As a result, the null 3. Chen K, Zhou F-Y, Yuan X-F (2019) Hybrid particle swarm
hypothesis of identical accuracy is rejected at the default optimization with spiral-shaped mechanism for feature selection.
Expert Sys Appl. https://doi.org/10.1016/j.eswa.2019.03.039
significance level of 5%. The f-ratio value resulted is 7.44. 4. Bellman RE (2015) Adaptive control processes. Princeton
5.62E - 06 is the p-value. At p \ 0.05, the result is sig- University Press, New Jersey
nificant (Table 11). 5. Al-Wajih R, Abdulkadir SJ, Aziz N, Al-Tashi Q, Talpur N (2021)
Hybrid binary grey wolf with harris hawks optimizer for feature
selection. IEEE Access 9:31662–31677
6. Dash M, Liu H (1997) Feature selection for classification. Intell
6 Conclusion Data Anal 1(1–4):131–156
7. Heidari AA, Mirjalili S, Faris H, Aljarah I, Mafarja M, Chen H
To improve the performance of the existing wrapper-based (2019) Harris hawks optimization: algorithm and applications.
Future Generat Comp Sys. https://doi.org/10.1016/j.future.2019.
BWOA algorithm, a hybrid metaheuristic technique was 02.028
developed. The new approach BWOAHHO, investigated 8. D. Karaboga (2005) ‘‘An idea based on honey bee swarm for
using the HHO algorithm instead of the exploration phase numerical optimization,’’
of native BWOA. Using 18 standard UCI datasets for 9. Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer.
Adv Eng Softw. https://doi.org/10.1016/j.advengsoft.2013.12.007
comparison. The effectiveness of the presented technique 10. Panwar LK, Reddy S, Verma A, Panigrahi BK, Kumar R (2018)
has been assessed and contrasted. The native BGA, BPSO, Binary grey wolf optimizer for large scale unit commitment
BWOA, BHHO, and WOASAT-2 feature selection algo- problem. Swarm Evol Comput 38:251–266
rithms were tested and compared to the new hybrid

123
Neural Computing and Applications

11. Mirjalili S (2015) The ant lion optimizer. Adv Eng Softw. https:// scheduling problems. Appl Soft Comput. https://doi.org/10.1016/
doi.org/10.1016/j.advengsoft.2015.01.010 j.asoc.2019.105954
12. Mirjalili S, Lewis A (2016) The whale optimization algorithm. 32. Hussain K, Neggaz N, Zhu W, Houssein EH (2021) An efficient
Adv Eng Softw. https://doi.org/10.1016/j.advengsoft.2016.01.008 hybrid sine-cosine Harris hawks optimization for low and high-
13. Thaher T, Heidari AA, Mafarja M, Dong JS, Mirjalili S (2020) dimensional feature selection. Expert Syst Appl 176:114778
Binary Harris Hawks optimizer for high-dimensional, low sample 33. Houssein EH, Hosney ME, Elhoseny M, Oliva D, Mohamed
size feature selection. Evolutionary machine learning techniques. WM, Hassaballah M (2020) Hybrid Harris hawks optimization
Springer, Berlin, pp 251–272 with cuckoo search for drug design and discovery in chemoin-
14. Chandrashekar G, Sahin F (2014) A survey on feature selection formatics. Sci Rep 10(1):1–22
methods. Comput Electr Eng. https://doi.org/10.1016/j.compelec 34. Reddy KS, Panwar LK, Panigrahi B, Kumar R (2018) A new
eng.2013.11.024 binary variant of sine–cosine algorithm: development and appli-
15. Liu H, Yu L (2005) Toward integrating feature selection algo- cation to solve profit-based unit commitment problem. Arab J Sci
rithms for classification and clustering. IEEE Trans Knowl Data Eng 43(8):4041–4056
Eng 17(4):491–502 35. Yıldız AR, Yıldız BS, Sait SM, Bureerat S, Pholdee N (2019) A
16. Muni DP, Pal NR, Das J (2006) ‘‘Genetic programming for new hybrid Harris hawks-Nelder-Mead optimization algorithm
simultaneous feature selection and classifier design,.’’ IEEE for solving design and manufacturing problems. Mater Test
Trans Sys, Man, Cybernet Part B (Cybernetics) 36(1):106–117 61(8):735–743. https://doi.org/10.3139/120.111378
17. Lin J-Y, Ke H-R, Chien B-C, Yang W-P (2008) Classifier design 36. Barshandeh S, Piri F, Sangani SR (2020) HMPA: an innovative
with feature selection and feature extraction using layered genetic hybrid multi-population algorithm based on artificial ecosystem-
programming. Expert Syst Appl 34(2):1384–1393 based and Harris Hawks optimization algorithms for engineering
18. Talbi EG (2002) A taxonomy of hybrid metaheuristics. J Heurist. problems. Eng Comput. https://doi.org/10.1007/s00366-020-
https://doi.org/10.1023/A:1016540724870 01120-w
19. Il-Seok O, Jin-Seon L, Byung-Ro M (2004) Hybrid genetic 37. Mafarja MM, Mirjalili S (2017) Hybrid whale optimization
algorithms for feature selection. IEEE Trans Pattern Anal Mach algorithm with simulated annealing for feature selection. Neu-
Intell 26(11):1424–1437. https://doi.org/10.1109/TPAMI.2004. rocomputing. https://doi.org/10.1016/j.neucom.2017.04.053
105 38. Emary E, Zawbaa HM, Hassanien AE (2016) Binary ant lion
20. Dong H, Li T, Ding R, Sun J (2018) A novel hybrid genetic approaches for feature selection. Neurocomputing. https://doi.
algorithm with granular information for feature selection and org/10.1016/j.neucom.2016.03.101
optimization. Appl Soft Comput. https://doi.org/10.1016/j.asoc. 39. Chuang L-Y, Chang H-W, Tu C-J, Yang C-H (2008) Improved
2017.12.048 binary PSO for feature selection using gene expression data.
21. Huang J, Cai Y, Xu X (2007) A hybrid genetic algorithm for Computat Biol Chem. https://doi.org/10.1016/j.compbiolchem.
feature selection wrapper based on mutual information. Patt 2007.09.005
Recogn Lett. https://doi.org/10.1016/j.patrec.2007.05.011 40. Gou J, Qiu W, Yi Z, Shen X, Zhan Y, Ou W (2019) Locality
22. Liu X-Y, Liang Y, Wang S, Yang Z-Y, Ye H-S (2018) A hybrid constrained representation-based K-nearest neighbor classifica-
genetic algorithm with wrapper-embedded approaches for feature tion. Knowl-Based Sys. https://doi.org/10.1016/j.knosys.2019.01.
selection. IEEE Access 6:22863–22874 016
23. Huang C-L, Dun J-F (2008) A distributed PSO–SVM hybrid 41. Gou J, Ma H, Ou W, Zeng S, Rao Y, Yang H (2019) A gener-
system with feature selection and parameter optimization. Appl alized mean distance-based k-nearest neighbor classifier. Expert
Soft Comput. https://doi.org/10.1016/j.asoc.2007.10.007 Syst Appl 115:356–372
24. Abdel-Basset M, Ding W, El-Shahat D (2021) A hybrid Harris 42. Yesilbudak M, Sagiroglu S, Colak I (2017) A novel implemen-
Hawks optimization algorithm with simulated annealing for tation of kNN classifier based on multi-tupled meteorological
feature selection. Artif Intell Rev 54(1):593–637 input data for wind power prediction. Energy Convers Manage.
25. Javidrad F, Nazari M, Javidrad HR (2018) Optimum stacking https://doi.org/10.1016/j.enconman.2016.12.094
sequence design of laminates using a hybrid PSO-SA method. 43. Asuncion A, Newman D (2007) ‘‘UCI machine learning reposi-
Comp Struct. https://doi.org/10.1016/j.compstruct.2017.11.074 tory,’’ ed: Irvine, CA, USA
26. Eappen G, Shankar T (2020) Hybrid PSO-GSA for energy effi- 44. Al-Tashi Q, Kadir SJA, Rais HM, Mirjalili S, Alhussian H (2019)
cient spectrum sensing in cognitive radio network. Phys Com- Binary optimization using hybrid grey wolf optimization for
mun. https://doi.org/10.1016/j.phycom.2020.101091 feature selection. IEEE Access 7:39496–39508. https://doi.org/
27. Abdel-Basset M, El-Shahat D, El-henawy I, De Albuquerque 10.1109/ACCESS.2019.2906757
Victor Hugo C., Mirjalili S (2020) A new fusion of grey wolf 45. Imandoust SB, Bolandraftar M (2013) Application of k-nearest
optimizer algorithm with a two-phase mutation for feature neighbor (knn) approach for predicting economic events: theo-
selection. Expert Sys Appl. https://doi.org/10.1016/j.eswa.2019. retical background. Int J Eng Res Appl 3(5):605–610
112824 46. Wang L, Khan L, Thuraisingham B (2008) ‘‘An effective evi-
28. Mafarja M, Qasem A, Heidari AA, Aljarah I, Faris H, Mirjalili S dence theory based k-nearest neighbor (knn) classification,’’ In
(2020) Efficient hybrid nature-inspired binary optimizers for 2008 IEEE/WIC/ACM International conference on web intelli-
feature selection. Cogn Comput 12(1):150–175 gence and intelligent agent technology, 2008, vol. 1: IEEE,
29. Mohammadzadeh H, Gharehchopogh FS (2021) A novel hybrid pp. 797–801
whale optimization algorithm with flower pollination algorithm 47. Al-wajih R, Abdulakaddir SJ, Aziz NBA, Al-tashi Q (2020)
for feature selection: Case study Email spam detection. Comput ‘‘Binary Grey Wolf Optimizer with K-Nearest Neighbor classifier
Intell 37(1):176–209 for Feature Selection,’’ In 2020 International conference on
30. Got A, Moussaoui A, Zouache D (2021) Hybrid filter-wrapper computational intelligence (ICCI), 8–9 Oct. 2020 2020,
feature selection using whale optimization algorithm: a multi- pp. 130–136, doi: https://doi.org/10.1109/ICCI51257.2020.
objective approach. Expert Sys Appl 183:115312 9247792
31. Liu M, Yao X, Li Y (2020) Hybrid whale optimization algorithm 48. Khanesar MA, Teshnehlab M, Shoorehdeli MA (2007) ‘‘A novel
enhanced with Lévy flight and differential evolution for job shop binary particle swarm optimization,’’ In: 2007 Mediterranean
conference on control & automation, IEEE, pp. 1–6

123
Neural Computing and Applications

49. Mafarja M, Aljarah I, Faris H, Hammouri AI, Ala’M A-Z, Mir- 51. Emary E, Zawbaa HM, Hassanien AE (2016) Binary grey wolf
jalili S (2019) Binary grasshopper optimisation algorithm optimization approaches for feature selection. Neurocomputing
approaches for feature selection problems. Expert Sys Appl 172:371–381
117:267–286
50. Abdulrauf Sharifai G, Zainol Z (2020) ‘‘Feature selection for Publisher’s Note Springer Nature remains neutral with regard to
high-dimensional and imbalanced biomedical data based on jurisdictional claims in published maps and institutional affiliations.
robust correlation based redundancy and binary grasshopper
optimization algorithm,’’ Genes, vol. 11, no. 7, p. 717, 2020.
[Online]. Available: https://www.mdpi.com/2073-4425/11/7/717

123

View publication stats

You might also like