Journal of Ambient Intelligence and Humanized Computing

[Link]

ORIGINAL RESEARCH

Improved salp swarm algorithm based on particle swarm optimization

for feature selection
Rehab Ali Ibrahim1 · Ahmed A. Ewees2,3 · Diego Oliva4 · Mohamed Abd Elaziz5 · Songfeng Lu1,6

Received: 22 January 2018 / Accepted: 3 September 2018

© Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract
Feature selection (FS) is a machine learning process commonly used to reduce the high dimensionality problems of datasets.
This task permits to extract the most representative information of high sized pools of data, reducing the computational
effort in other tasks as classification. This article presents a hybrid optimization method for the FS problem; it combines
the slap swarm algorithm (SSA) with the particle swarm optimization. The hybridization between both approaches creates
an algorithm called SSAPSO, in which the efficacy of the exploration and the exploitation steps is improved. To verify
the performance of the proposed algorithm, it is tested over two experimental series, in the first one, it is compared with
other similar approaches using benchmark functions. Meanwhile, in the second set of experiments, the SSAPSO is used to
determine the best set of features using different UCI datasets. Where the redundant or the confusing features are removed
from the original dataset while keeping or yielding a better accuracy. The experimental results provide the evidence of the
enhancement in the SSAPSO regarding the performance and the accuracy without affecting the computational effort.

Keywords Salp swarm algorithm · Particle swarm optimization · Feature selection · Global optimization · Swarm
techniques

1 Introduction
* Songfeng Lu
lusongfeng@[Link]
Feature selection (FS) has become the focus of many types
Rehab Ali Ibrahim
rehab100r@[Link] of researches for machine learning and data mining in many
areas of applications for datasets of tens or hundreds or
Ahmed A. Ewees
[Link]@[Link] even thousands of variables (Guyon and Elisseeff 2003). FS
has been widely applied in several areas like text process-
Diego Oliva
[Link]@[Link] ing of internet documents (Al-Ayyoub et al. 2017; Chikh
and Chikhi 2017; Saravanan and Rajesh Babu 2017), gene
Mohamed Abd Elaziz
abd_el_aziz_m@[Link] expression array analysis (Li and Wong 2002), text catego-
rization (Yang et al. 2002), genomics (Kohane et al. 2002),
1
School of Computer Science and Technology, Huazhong cancer detection (Prabukumar et al. 2017), image classifica-
University of Science and Technology, Wuhan 430074, tion (Ibrahim et al. 2018), computer vision (Arigbabu et al.
China
2016), signal processing (Kung et al. 2010), bioinformatics
2
University of Bisha, Bisha, Kingdom of Saudi Arabia (Awada et al. 2012), image retrieval (El Aziz et al. 2018a;
3
Department of Computer, Damietta University, Damietta, Li and Wang 2015; Wang et al. 2018), medical applications
Egypt (Chang et al. 2012), combinatorial chemistry (Jensen et al.
4
Departamento de Ciencias Computacionales, Universidad de 2009) and others.
Guadalajara, CUCEI, Av. Revolucion 1500, Guadalajara, Jal, Feature selection is of three benefits: (1) improving the
Mexico
prediction performance of the predictors, (2) providing a
5
Department of Mathematics, Faculty of Science, Zagazig better understanding of the underlying process that generates
University, Zagazig, Egypt
the data fastly (3) has more cost-effective predictors (Han
6
Shenzhen Huazhong University of Science and Technology et al. 2011). These contributions cover a wide area of aspects
Research Institute, Shenzhen 518063, China

13
Vol.:(0123456789)
 R. A. Ibrahim et al.

of many problems: providing a good definition of the fitness up of a population of artificial agents inspired by the social
function, feature construction, multivariate feature selection, behavior of animals (fish, birds, fireflies, etc.) in the real
feature ranking, efficient search methods, and feature validity world. For example, authors in (Chen et al. 2014; Tanaka
assessment methods. Thus, the top reasons to use the feature et al. 2007) proposed using the binary particle swarm opti-
selection are that it enables the machine learning algorithm mization (PSO) for FS. Zhang in (Zhang and Sun 2002)
to train faster, reduces the complexity of a model and makes applied the Tabu search in FS. Artificial fish swarm (AFS)
it easier to interpret, improves the accuracy of a model if the algorithm (Li and Wong 2002) mimics the stimulant reac-
right subset is chosen, and reduces the overfitting. tion by controlling the tail and fin. AFS is a robust stochastic
In general, feature selection is a difficult task due to a technique based on the fish movement and its intelligence
large search space, where the total number of possible solu- during the food finding process. Artificial bee colony (ABC)
tions is 2l for a dataset of l features. The task becomes more (Karaboga and Basturk 2007) is a numerical optimization
challenge as l is increasing in many areas with the advances algorithm based on the foraging behavior of honey bees.
in the data collection techniques and the high complexity Many other recent swarms as social spider optimization
of the problems. A variety of search techniques have been (SSO) (Cuevas and Cienfuegos 2014), ant colony optimiza-
applied to feature selection, such as complete search, greedy tion (ACO) (Moradi and Rostami 2015), gray wolf optimiza-
search, heuristic search (Zhong 2001), and random search tion (GWO) (Mirjalili et al. 2014), and salp swarm algorithm
(Lai et al. 2006; Raymer et al. 2000). (SSA) (Mirjalili et al. 2017) have also been proposed for
Several meta-heuristic techniques mimic the biological complex optimization problems. However, most existing FS
and the physical behavior in nature, and it was proven that methods as the previously mentioned still suffer from trap-
they are strong methods for global optimization. Genetic ping in local optima and high computational cost (Liu et al.
algorithm (GA) (Goldberg and Holland 1988) was the first 2011; Unler and Murat 2010). Therefore, recent literature
evolutionary algorithm introduced in this field and has been combined two swarms for strong global optimization, where
developed based on the natural process of evolution through this adds more flexibility to the basic swarm in exploring the
reproduction. GA can solve complex and non-linear prob- population and ensures the diversity of it.
lems, but it has some drawbacks as that its performance is Many works in this context, in (Yeh et al. 2016), the
low and may be stuck in local optima. authors proposed a hybrid method which can spot a better
In recent years, a big amount of feature selection meth- optimal features subset. Which it is a hybrid of a swarm
ods have been proposed (Mirjalili et al. 2014), in which, called simplified swarm optimization which is a soft
there are two key issues in the structure of a feature selection computing algorithm evolved from the PSO algorithm
method: search strategies and evaluating measures. Con- and is developed to overcome the drawback of PSO in
cerning search strategies, complete, heuristic, and random the discrete problem. Meanwhile, the information gain
strategies were proposed as stated before. And, with respect as a filter selects the most informative features from the
to evaluating measures, these methods can be divided into original dataset and the support vector machine serves as
two classes: classification (Gasca 2006; Kohavi 1994; Neu- a classifier to evaluate the performance. However, this
mann et al. 2005) and classification independent (Dash method suffers from the problem of the premature con-
and Liu 2003; Kira and Rendell 1992). The first approach vergence. In (Hafez et al. 2015), it is presented a system
employs a learning algorithm to evaluate the quality of the for the FS based on a combination of Monkey Algorithm
selected features based on the classification accuracies or (MA) with Krill Herd Algorithm (KHA). In (Yamuna and
the contribution to the classification boundary, such as the Thamaraichelvi 2016), the authors presented a hybridiza-
so-called wrapper method (Kohavi 1994) and weight-based tion of the particle swarm optimization and the firefly
algorithms (Guyon et al. 2002; Xie et al. 2006). While, the (PSO-FF) algorithms for FS to diagnose the childhood
latter constructs a classifier independent measure to evaluate typical teratoid/rhabdoid tumor (AT/RT) in magnetic res-
the importance of the features, such as interclass distance onance brain images and hemochromatosis in computed
(Kira and Rendell 1992), mutual information (Yao 2003), tomography (CT) liver images. The authors in (Menghour
dependence measure (Modrzejewski 1993), and consistency and Souici-Meslati 2016) proposed a hybrid bio-inspired
measure (Dash and Liu 2003). approach applied to the FS problem, which is based on
Considering the above, different methods have been two swarm intelligence methods: ant colony optimization
applied to FS; whereas, the best practice is to use the opti- (ACO) and particle swarm optimization (PSO). Also, in
mization technique to improve the effectiveness of these (Chuang et al. 2009), it is proposed to combine the Tabu
methods and assist in escaping from getting stuck in local search (TS) and the binary particle swarm optimization
targets. Evolutionary and swarm intelligence methods model (BPSO) for FS. BPSO acts as a local optimizer, each time
the FS as an optimization problem. Swarm intelligence is a the TS has been run for a single generation, this is for
computational intelligence-based approach which is made the gene expression in the classification of cancer, but

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

this approach depends on the fewest number of features 2 Preliminaries

and this may not represent the whole dataset, and thus
may stick in local points. The authors of (Thangavel and 2.1 Salp swarm algorithm (SSA)
Velayutham 2011) introduced a hybrid of ant colony
optimization, bee colony optimization, fuzzy C-means Salp swarm algorithm (SSA) is a new optimization technique
and genetic algorithm for the selection of the features (Mirjalili et al. 2017), produced to solve numerous kinds of
extracted from the mammogram image. (Karnan et al. optimization problems. It mimics the behavior of Salps in
2006) presented a hybrid of spatial gray level dependence nature; salps are a kind from the Salpidae’s family, and they
method (SGLDM), ant colony optimization, and genetic are barrel-shaped planktonic tunicate. In addition, they are
algorithm for the selection and the classification of fea- similar to jellyfishes in tissues, and moving behavior and
tures extracted from the mammogram image. Features are their weights have a high water percentage (Henschke et al.
extracted using SGLDM. Feature selection is performed 2016). They move through contracting, so pumping water
using GA and ACO, which then is fed to a three-layer through their jellied bodies change their positions (Madin
BPN hybrid with ACO for classification, and many other 1990). Salps, in oceans, have a swarm behavior called salp
hybrids for several applications (El Aziz et al. 2016, 2017, chain, this behavior may help salps in foraging and doing
2018b; Elaziz et al. 2017; Ewees et al. 2017a, b, 2018; better movement using fast harmonious changes (Anderson
Ibrahim et al. 2017). and Bone 1980; Sutherland and Weihs 2017). Based on this
In the same context, a new swarm called salp swarm conduct, the authors of (Mirjalili et al. 2017) modeled the
algorithm (SSA) (Mirjalili et al. 2017) is introduced, salp chains in a mathematical form and tested it in optimiza-
which mimics the behavior of salps in nature, where salps tion problems.
are a kind from the Salpidae’s family, and they are bar- SSA starts by dividing the population into two groups
rel-shaped planktonic tunicate. However, the SSA suffers namely the leader and the followers. The front salp of the
from the same limitations that face the other metaheuris- chain is called the leader, and the other salps are called the
tic (MH) algorithms such as slow convergence since its followers. The salps’ position is determined in n-dimensions
ability exploitation is not good. Therefore, to avoid this which represent the search space of a problem and n represents
limitation, in this paper, it is combined with the particle the problem’s variables. These salps search for a food source
swarm optimization (PSO) to make an alternative hybrid which indicates the target of the swarm. The position should
swarm algorithm. Where, this integration adds more flex- be updated frequently, so, the following equation is used to
ibility to the SSA to improve its ability exploitation and perform this action to the salp leader:
ensure the diversity of it, as well as reaches the optimal {
value quickly. 1 Fj + c1 ((ubj − lbj ) × c2 + lbj ) c3 ≤ 0
xj =
Fj − c1 ((ubj − lbj ) × c2 + lbj ) c3 > 0 (1)
In general, the proposed algorithm starts by generat-
ing a random population then computing the objective
function for each solution. Then the best solution is deter- where xj1 is the position of the leader within j-th dimension,
mined, and the probability of each objective function is where the food source in this dimension is Fj , the upper and
computed since it is used to update the current solution the lower bounds are ubj and lbj , respectively. c2 and c3 are
through switching between the SSA and the PSO and this generated randomly in the range [0, 1] to maintain the search
leads to a balance between the exploration and the exploi- space. Also, the parameter c1 is the very important coeffi-
tation of the SSA. The proposed algorithm is compared cient in this algorithm, due to its role in the balancing
with other algorithms through a set of two experimental between the exploration phase and the exploitation phase
series, in the first series, the proposed algorithm is com- and it is calculated as follows:
pared with the standard SSA algorithm to solve a set of ( )2
(2)
4t
15 benchmark functions to find the global solution. Mean- c1 = 2e
− tmax ,
while, in the second experimental series, the proposed
algorithm is compared with the other four algorithms to where t and tmax indicate the current iteration and the max
select the optimal subset of features to improve the clas- iterations’ number, respectively. After updating the leader’s
sification accuracy of some UCI datasets. position, the SSA starts to update the followers’ position
This paper is organized as follows, Section 2 explains using the following equation:
the main principals of the Salp Swarm Algorithm and the (
1 i
)
Particle Swarm Optimization. Section 3 describes the xji = xj + xji−1 (3)
2
proposed algorithm. The experiments and discussion are
given in Section 4; whereas, the last section includes the
conclusion and the future work.

13
 R. A. Ibrahim et al.

xji is the i-th follower position within j-th dimension and This sequence is repeated until meeting the stopping cri-
i is greater than 1. The final steps of the SSA are given in teria (e.g. fixed numbers of iterations). The final steps of the
Algorithm 1. PSO are given in Algorithm 2.

Algorithm 1 Salp Swarm Algorithm (SSA)

1: Initialize a population X. Algorithm 2 Particle Swarm Optimization Algorithm
2: repeat (PSO)
3: Compute the objective function for each solution xi . 1: Initialize the population.
4: Update the best salp (solution) (F = X b ). 2: repeat
5: Update c1 using Eq. (2). 3: for i = 1 to Population Size do
6: for i = 1 : N do 4: Calculate the objective function value.
7: if i == 1 then 5: If the fitness value is better than the best
8: Update the position of salp using Eq. (1) one(P Best).
9: else 6: Then set the current value as the new P Best.
10: Update the position of salp using Eq. (3) 7: Choose the particle with the best fitness value of all
11: end if the particles as GBest .
12: end for 8: Calculate a new velocity in accordance with Eq.(5).
13: until (t < tmax ) 9: Update the particle position in accordance with
14: Return the best solution F . Eq.(4).
10: end for
2.2 Particle swarm optimization 11: until termination criterion is met

Particle swarm optimization (PSO) was introduced by Eber- 3 The proposed algorithm
hart and Kennedy (1995), it mimics the evolvement of the
knowledge of a social behavior (Noman et al. 2009) and simu- In this section, the structure of the proposed algorithm is
lates the group communication behavior when sharing private explained. It is called SSAPSO which combines the SSA and
knowledge concerning migrating, flocking, or hunting. This the PSO algorithms. The basic structure of the SSA algo-
group and its members represent a solution and are called a rithm is modified by improving the updating phase of the
swarm and particles, respectively (Niknam and Amiri 2010). population’s position. This modification merges the update
A particle relies on its experience and neighbors to update its mechanism of the PSO into the main structure of the SSA.
position. This integration adds more flexibility to the SSA in explor-
The swarm begins by creating a set of random particles ing the population and ensures the diversity of it, as well as
as well as generating their positions ( xi) and velocity (vi) in a reaches the optimal value quickly.
dimension j-th. After that, PSO starts its main loop to evalu- In general, the main structure of the proposed SSAPSO
ate each particle by performing a fitness function; the result is algorithm is given in Fig. 1. Where first step in the pro-
checked with its best and global best values. The Eqs. (4) and posed SSAPSO is to define the parameters and generate the
(5) represent the method that is used to update the particles population which represents a set of solutions for the given
positions (Niknam and Amiri 2010; Noman et al. 2009). problem (feature selection). Then the performance of each
solution is evaluated by computing the fitness function for
xij(t+1) =xij(t) + v(t+1) (4) each one and determine the best of them. The next step in
ij
the proposed SSAPSO algorithm is to update the current
population by using either the SSA or PSO algorithm which
v(t+1)
ij
=wv(t)
ij
p(t) g(t)
+ c1 r1 (xij − xij(t) ) + c2 r2 (xj − xij(t) ) (5) depends on the quality of the fitness function (measured by
its probability as will discuss later). In which if the probabil-
where xij is the i-th particle position in the j-th dimension, ity of fitness function, for the current solution, is greater than
vij is the i-th velocity in the j-th dimension, t defines the cur- 0.5 then the SSA, otherwise, the PSO is used. Thereafter, the
rent iteration, w indicates an inertia weight and applied to fitness function for each solution is computed and the best
improve the speed of the population convergence. The con- solution is determined after updating the population. The
stants c1 and c2 are acceleration coefficients. xij defines the
p(t)
next step is to check if the stop conditions are satisfied then
best previous position of particle i in j-th dimension; xj
g(t)
return by the best solution, otherwise, repeat the previous
defines the global best position in j-th dimension. r1 and r2 steps from computing the probability to the end. These steps
are random parameters ∈ [0, 1]. are discussed with more details, in the following.
The SSAPSO algorithm starts by defining the initial val-
ues of the SSA and the PSO, then SSA generates a random

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

other elements are ignored since they represent the irrelevant

features). The next step is to compute the objective function
for each xi as in Eq. (7):
( )
( ) |xi (t)|
f xi (t) = 𝜉Exi (t) + (1 − 𝜉) , (7)
|C|

where Exi (t) represents the error of the classification per-

formed by the effective classifier, K-NN classifier; while
the second term represents the number of selected features.
In order to balance between the classification error and the
number of the selected features, the parameter 𝜉 ∈ [0, 1] is
used. The next step is to compute the probability of each
fitness function ( Proi ) as:

f
Proi = ∑Ni (8)
f
i=1 i

According to the Proi value, the current solution xi will be

updated using the SSA or the PSO. For example, Proi > 0.5,
then the SSA is used as explained in Algorithm 1, otherwise,
the PSO algorithm as discussed in Algorithm 2.
The fitness function is computed for each updated solu-
tion, and the best solution is updated. This sequence is
iterated until meeting the stopping condition (the proposed
SSAPSO algorithm applies for the max iteration’s number
as a stopping condition).

4 Experiments study

In this section, we apply two experiments to evaluate the

quality of the proposed algorithm. The first one uses 15
well-known benchmark functions, and the results are com-
pared with the standard version of the SSA.
The second experiment uses ten common datasets to
test the proposed algorithm in the features selection prob-
lem, and the results are compared with four algorithms
namely SSA, PSO, GA, and BAT.
Fig. 1  The main structure of the proposed algorithm
4.1 Performance measures
population X of size N in dimension D, then SSA calculates In order to evaluate the performance of the proposed
the food fitness for each solution xi , i = 1, 2, .., N . However, algorithm, a set of performance measures are used in this
before computing the objective function, each solution xi is paper. The definitions of these measures are given as in
converted to a binary vector (that contains only 1’s and 0’s) the following:
according to the value of a random threshold 𝜖 ∈ [0, 1] using
the following equation: – Mean value It is an average of the beat obtained values
{ over several runs, it is calculated as:
1 if 1+e1−xi (t) > 𝜖
xi (t + 1) = (6) Nr
0 otherwise 1 ∑ ∗,
𝜇F = F (9)
Nr i=1 i
Therefore, only the xi elements that are corresponding to
1’s are chosen to represent the selected features (also, the

13
 R. A. Ibrahim et al.

where Nr is the number of runs and Fi∗ is the best Also, Ns represent the number of samples in the given
obtained value. dataset
– Maximum value It indicates the maximum obtained – Selection ratio is defined as in the following equation:
value over different runs.
Nr i
1 ∑ |xbest |
Max = min F ∗ SelR = , (16)
1≤i≤Nr i (10) M i=1 n

– Minimum value It indicates the minimum obtained where |xbest

i
| and n are the total number of the selected
value over different runs. features at the ith run, and the total number of features
of the dataset, respectively.
Min = max Fi∗ (11)
1≤i≤Nr
4.2 Parameters setting
– Standard deviation It is used to indicate if the algo-
rithm can produce the best value over different runs or
The proposed algorithm is compared with the other four
not.
algorithms namely, particle swarm optimization (PSO)
√
√
√ 1 ∑ Nr (Inbarani et al. 2014), salp swarm algorithm (SSA) (Mir-
STDF = √ (F − 𝜇F )2 (12) jalili et al. 2017), genetic algorithm (Goldberg and Holland
Nr − 1 i=1 i 1988), bat algorithm (Rodrigues et al. 2014). The parameters
of each algorithm are given in Table 1. These comparative
– Success rate It is calculated to check if the algorithm algorithms are implemented in Matlab2016 that is installed
reaches a predefined value called (value-to-reach over the Windows environment with 64-bit support.
(VTR)). The environment setting of the experiment series 1 (i.e.
(
Number of times reached VTR
) benchmark function) are population = 30, iterations = 200,
SR = × 100 (13) value to reach = 10e−5, and the lower and the upper bound
Nr
are set as shown in Table 2 for the global optimization.
– Acceleration rate It calculates the number of function Whereas, the environment setting of the experiment
calls (NFC) of the algorithm. series 2 (i.e. feature selection) are the population = 30, lower
( ) bound = 0, upper bound = 1, iterations = 100. In addition,
NFCOBLGOA
AR = 1 − × 100 (14) each dataset is divided into two groups (, i.e., training group
NFCGOA and testing group) in each run using 5-fold cross-validation
– Time average To check the algorithm time complexity. to ensure the variety of the samples in each run. To assess
– Root mean square error (RMSE) It is a computed error the accuracy of the algorithms, the K-Nearest Neighbour
between the predicted and the original values, in order (KNN) classifier is used.
to evaluate the performance of the selected features
through using a classifier. 4.3 Experiment series 1: benchmark functions
√
√
√1 ∑ Ns This experiment compares the results of the proposed algo-
RMSE = √ (P − Ti )2 ), (15) rithm against the standard version of the SSA over 15 bench-
Ns i=1 i
mark functions. Table 2 shows the details of these functions
that vary between the unimodal and the multimodel type.
where P is the estimated fitness function values from an
algorithm and T is the original fitness function values.

Table 1  The parameters setting of all algorithms

Algorithm Parameters setting

SSA c2 and c3 are random values ∈ [1, 0]

PSO wMax = 0.9;wMin = 0.2;C1 = 2;C2 = 2
SSAPSO It uses the parameters of SSA and PSO.
GA Crossover type is 1; crossover probability = 1; mutation probability = 0.01
BAT loudness = 0.5; pulse rate = 0.5; frequency minimum = 0; frequency maximum = 2

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

Table 2  Set of benchmark optimization functions

ID Equation Lower Upper Dimension type
∑n
F1 f (x) = xi2 − 100 100 10 Unimodal
∑i=1
F2 f (x) =
n n
i=1 �xi � + Πi=1 �xi �
− 10 10 10 Unimodal
F3 ∑n ∑i 2 − 100 100 10 Unimodal
f (x) = i=1 ( j−1 xi )
F4 f (x) = maxi {|xi |, 1 ⩽ i ⩽ n} − 100 100 10 Unimodal
F5 f (x) = ∑n−1 [100(xi+1 − x2 )2 + (xi − 1)2 ] − 30 30 10 Unimodal
∑i=1 i
F6 f (x) = ni=1 ([xi + 0.5])2 − 100 100 10 Unimodal
∑
F7 f (x) = ni=1 ixi4 + random[0, 1] − 1.28 1.28 10 Unimodal
∑ √
F8 f (x) = n −xi sin( �xi �) − 500 500 10 Multimodal
i=1
∑
F9 f (x) = ni=1 [xi2 − 10cos(2𝜋xi ) + 10] − 5.12 5.12 10 Multimodal
�
F10 f (x) = −20exp(−0.2 1 ∑n x2 ) − exp( 1 ∑n cos(2𝜋x )) + 20 + e − 32 32 10 Multimodal
n i=1 i n i=1 i

F11 f (x) = 1 ∑n 2 n x − 600 600 10 Multimodal

√i
4000 i=1 xi − Πi=1 cos( i ) + 1
∑ ∑n
F12 f (x) = 𝜋
{10sin2 (𝜋y1 ) + n−1 2 2 − 1)2 } + − 50 50 10 Multimodal
n i=1 (yi − 1) [1 + 10sin (𝜋yi+1 )] + (yn i=1 u(xi , 10, 100, 4)
⎧ k(x − a)m , x > a
⎪ i i
u(xi , a, k, m) = ⎨ 0, − a ⩽ xi ⩽ a
⎪ k(−xi − a)m , xi < −a
⎩
∑
F13 f (x) = 0.1{sin2 (3𝜋x1 ) + ni=1 (xi − 1)2 [1 + sin2 (3𝜋xi + 1)] + (xn − 1)2 [1 + sin2 (2𝜋xn )]} − 50 50 10 Multimodal
∑n
+ i=1 u(xi , 5, 100, 4)
F14 f (x) = ( 1 + ∑25 ∑2 1 )− 1 − 65.536 65.536 2 Multimodal
500 j=1 j+ 6
i=1 (xi −aij )

F15 f (x) = (∑11 [a − x1 (b2i +bi x2 ) 2

] −5 5 4 Multimodal
i=1 i b2i +bi x3 +x4

Table 3  The statistical results F. No SSA SSAPSO

of the SSA and the SSAPSO
algorithms using 15 benchmark Mean Min Max STD Mean Min Max STD
functions
F1 2.81E−09 1.09E−09 6.67E−09 1.56E−09 8.17E−10 3.38E−31 2.29E−09 6.82E−10
F2 6.19E−02 1.31E−05 5.78E−01 1.29E−01 8.21E−23 5.02E−31 1.44E−21 2.80E−22
F3 4.439 1.67E−02 42.5331 8.9681 1.09E−34 4.43E−48 1.69E−33 4.09E−34
F4 1.79E−03 2.51E−05 9.16E−03 2.65E−03 8.77E−21 4.17E−25 1.58E−19 3.16E−20
F5 10.9319 4.5011 76.935 13.0722 7.8707 7.3333 8.1156 2.62E−01
F6 2.84E−09 8.76E−10 7.25E−09 1.65E−09 6.40E−10 2.55E−11 1.38E−09 3.76E−10
F7 3.71E−02 4.72E−03 9.10E−02 2.08E−02 9.12E−04 5.97E−05 3.90E−03 8.48E−04
F8 − 2727.8 − 3378.95 − 2063.36 3.55E+02 − 2480.9 − 2689.49 − 2126.52 1.74E+02
F9 7.4107 3.9798 13.9294 3.1304 1.47E−14 0 1.56E−13 3.43E−14
F10 2.21E−05 9.65E−06 6.18E−05 1.15E−05 1.62E−15 8.88E−16 4.44E−15 1.46E−15
F11 1.02E−01 2.21E−02 4.53E−01 7.69E−02 3.30E−13 0 4.66E−12 9.35E−13
F12 1.4462 3.16E−02 3.0485 8.62E−01 1.50E−01 6.21E−02 2.54E−01 4.26E−02
F13 1.59E−04 1.71E−07 4.76E−04 1.56E−04 1.76E−05 2.38E−07 5.73E−05 1.50E−05
F14 1.4768 9.98E−01 3.9683 9.03E−01 9.98E−01 9.98E−01 9.98E−01 2.78E−16
F15 2.64E−03 3.83E−04 2.10E−02 5.09E−03 7.46E−04 3.84E−04 1.38E−03 2.55E−04

4.3.1 Results and discussion computed. These measures are mean, best, worst, standard
deviation, success rate, and the number of function calls.
The comparison results between the proposed algorithm and From Table 3, it can be observed that the proposed algo-
the SSA are given in Tables 3, 4, where each function is run rithm has better performance than the other algorithms based
30 times and the average of the performance measures are on the performance measure of the fitness function (mean,

13
 R. A. Ibrahim et al.

Table 4  The time, SR, and NFC F. No SSA SSAPSO AR

results of using 15 benchmark
functions Time SR NFC Time SR NFC

F1 0.4672 29 4808 0.6871 44 4278 1.1239

F2 0.4783 14 5873 0.7504 62 2623 2.2390
F3 0.6868 0 5970 0.9494 62 2549 2.3421
F4 0.4946 20 5923 0.7748 61 2759 2.1468
F5 0.5227 0 5970 0.808 0 5970 1
F6 0.5222 29 4819 0.6511 46 4212 1.1441
F7 0.5385 0 5970 0.8271 53 5919 1.0086
F8 0.5315 100 0 0.7555 100 0 5970
F9 0.5479 0 5970 0.7787 65 2270 2.6300
F10 0.5547 21 5678 0.8246 63 2667 2.1290
F11 0.5973 0 5970 0.7364 78 1684 3.5451
F12 0.7264 0 5970 0.9972 0 5970 1
F13 0.8665 34 5416 0.9895 63 3766 1.4381
F14 1.5093 0 5970 1.6649 0 5970 1
F15 0.535 81 4648 0.6428 92 3816 1
Avg. 22.2 52.60 1.6498

best, worst and std). In details, the SSAPSO had the lowest phases and the hybridization between the SSA and the
values in the mean measure, especially in the functions of PSO improves the quality of the basic SSA in solving the
the numbers 2, 3, 4, 9, 10, and 11; whereas the standard benchmark functions.
deviation proved that the SSAPSO’s results can reach the
same value in different runs.
From Table 4, we can see that the SSAPSO achieved 4.4 Experiment series 2: feature selection
very good results regarding all measures over 30 inde-
pendent runs against the SSA except for the time meas- In this section, the performance of the proposed algorithm
ure, however, the time of the proposed algorithm is little is tested over a set of UCI datasets that contains ten data-
higher than the SSA. The reason of the poor results in time sets. Also, the proposed algorithm is compared with the
measure of the SSAPSO is due to the PSO, in general, other four algorithms namely, SSA, GA, PSO, and BAT.
consumes more time to complete the exploration phase This section is organized, where the description of the ten
than the SSA; but this issue can be resolved if we added datasets is given followed by the parameter setting of each
a stopping condition as we mentioned above. The results algorithm, then the results and the discussion are given at
of the NFS also showed promising results of SSAPSO. In the end of this section.
this context, the SSAPSO called the function only 3375,
whereas the SSA called the function 4953.8 times. The 4.4.1 Datasets description
average AR (SSA versus SSAPSO) is 1.6498 which indi-
cates that the proposed SSAPSO is faster than the tradi- Table 5 shows the description of the ten datasets that are
tional SSA. In terms of the success rate, the SSAPSO had used to evaluate the performance of the proposed algorithm
outperformed SSA, where the average of the SR of the to select the optimal subset of features. These datasets have
proposed algorithm is higher 30% than the traditional SSA different characteristics such as a different number of sam-
algorithm overall the tested functions. ples, features, and classes.
Moreover, Fig. 2 shows the convergence curve of the
proposed SSAPSO and the SSA algorithms along the 4.4.2 Results and discussion
tested functions at the first of 200 iterations, from this
figure, it can be observed that the proposed SSAPSO can The comparison results of the proposed SSAPSO algorithm
be more convergent, and faster to the optimal solution than with the other algorithms are listed in Table 6 which shows
the SSA that requires a long time to reach. the average of the CPU time(s), the RMSE and the selec-
These results indicate that the SSAPSO has a good abil- tion ratio.
ity to balance between the exploration and the exploitation

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

Fig. 2  The convergence curve of the functions using the SSA and the SSAPSO algorithms

13
 R. A. Ibrahim et al.

Table 5  The datasets description RMSE value in the case of the datasets of WBDC, Iono-
No. Dataset Features Samples
sphere, Breastcancer, Lymphography, Clean1, and Clean2;
whereas PSO had the lowest values in the other datasets. So,
1 WBDC 31 569 SSAPSO can save the computational times and achieve the
2 Ionosphere 34 351 lowest RMSE in the large datasets than the other algorithms.
3 Hepatitis 19 155 SSAPSO also achieves the best reduction of the features in
4 Heart 13 270 all the datasets except for the Breastcancer dataset, it came
5 Breastcancer 9 683 in the second rank after SSA.
6 Sonar 60 208 In order to investigate the performance of the SSAPSO,
7 Lymphography 19 148 the selected features are passed to the KNN classifier to
8 Clean1 166 476 check if these features can increase the accuracy of the clas-
9 Clean2 166 6598 sification or not. So, in Table 7, we recorded the results of
10 Waveform 21 5000 this stage. From Table 7, it can be seen that, according to
the Clean2, WBDC, Ionosphere, Hepatitis, Lymphography,
Clean1 datasets results, the SSAPSO achieves the highest
From Table 6, it can be concluded that, in terms of the performance in both measures (i.e., accuracy and F-meas-
CPU time(s), SSAPSO had the lowest computation time ure). Where, for the Clean2 dataset, the SSAPSO is in the
in the case of the datasets of WBDC, Breastcancer, Lym- first rank followed by PSO, GA, SSA then BAT regarding the
phography, Clean1, and Waveform, also it has the lowest F-measure, and SSAPSO in the first rank followed by PSO,

Table 6  The results of the Dataset Measure SSA SSAPSO GA PSO BAT
selected features using ten
common datasets WBDC Time 38.69 38.68 46.42 39.37 43.57
RMSE 0.159 0.141 0.310 0.262 0.323
Selection 51 48 52 59 76
Ionosphere Time 39.62 36.69 40.46 36.51 36.83
RMSE 0.244 0.201 0.253 0.219 0.308
Selection 42 36 50 47 62
Hepatitis Time 37.47 36.31 39.43 35.65 36.35
RMSE 0.496 0.480 0.435 0.383 0.513
Selection 47 36 46 51 64
Heart Time 40.13 39.97 37.92 35.47 36.13
RMSE 0.407 0.39 0.392 0.347 0.488
Selection 57 55 59 60 78
Breastcancer Time 39.45 39.40 46.96 42 44.09
RMSE 0.185 0.141 0.143 0.142 0.163
Selection 54 56 57 61 84
Sonar Time 38.75 37.77 39.25 35.45 35.75
RMSE 0.227 0.117 0.090 0.087 0.1187
Selection 48 40 50 60 89
Lymphography Time 33.807 28.057 39.629 35.886 35.973
RMSE 0.313 0.310 0.384 0.348 0.420
Selection 55 48 56 63 74
Clean1 Time 45.81 44.43 50.31 59.74 48.88
RMSE 0.301 0.255 0.297 0.256 0.328
Selection 50 29 49 57 71
Clean2 Time 1855.2 1437.7 1022.1 1133.5 1138.1
RMSE 0.183 0.168 0.171 0.170 0.183
Selection 49 30 50 58 60
Waveform Time 213.8 210.7 239.7 246.37 275.4
RMSE 0.657 0.644 0.657 0.635 0.651
Selection 65 65 68 84 95

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

Table 7  The results of the Dataset Measure SSA SSAPSO GA PSO BAT
classifier to evaluate the
selected features over all WBDC Accuracy 0.974 0.980 0.904 0.932 0.895
datasets
F-measure 0.966 0.973 0.863 0.893 0.849
Ionosphere Accuracy 0.939 0.951 0.916 0.949 0.904
F-measure 0.955 0.962 0.936 0.959 0.930
Hepatitis Accuracy 0.717 0.748 0.643 0.721 0.629
F-measure 0.772 0.780 0.718 0.758 0.709
Heart Accuracy 0.823 0.847 0.790 0.831 0.761
F-measure 0.870 0.866 0.810 0.835 0.792
Breastcancer Accuracy 0.963 0.978 0.975 0.976 0.972
F-measure 0.9720 0.9834 0.9827 0.9833 0.9812
Sonar Accuracy 0.9443 0.9620 0.9670 0.9694 0.9527
F-measure 0.9399 0.9598 0.9637 0.9664 0.9481
Lymphography Accuracy 0.8469 0.9004 0.8563 0.8783 0.8229
F-measure 0.8942 0.9152 0.8562 0.8776 0.8213
Clean1 Accuracy 0.9088 0.9336 0.9107 0.9326 0.8900
F-measure 0.9144 0.9392 0.899 0.9384 0.8971
Clean2 Accuracy 0.9662 0.9831 0.9707 0.9711 0.9664
F-measure 0.9800 0.9831 0.9825 0.9829 0.9799
Waveform Accuracy 0.7849 0.7913 0.7838 0.7947 0.7929
F-measure 0.7847 0.7910 0.7836 0.7944 0.7928

GA, SSA, and BAT regarding the accuracy. For WBDC dataset, SSAPSO in the first followed by PSO, SSA, BAT,
dataset, SSAPSO is in the first rank followed by SSA, PSO, and GA regarding the F-measure and according to the accu-
GA, and BAT regarding the F-measure; while regarding the racy, SSAPSO then PSO, GA, and SSA, while the BAT had
accuracy, SSAPSO is in the first then SSA, PSO, GA, BAT. the last rank. SSAPSO, also came in the first rank of all the
For the Ionosphere dataset, SSAPSO is in the first followed remaining datasets, regarding accuracy followed by PSO,
by PSO, SSA, GA, and BAT in terms of the F-measure, and SSA, GA, and BAT, this is according to the average of them.
the same order according to the accuracy. For the Hepatitis While, in terms of the F-measure, SSAPSO was the best
dataset, SSAPSO then SSA, PSO, GA, and BAT in-terms of except for the Heart and the Sonar datasets, it came in the
the F-measure, and according to the accuracy, SSAPSO then second rank after PSO and SSA.
PSO, SSA, GA, and BAT. For the Lymphography dataset, In addition to as in Fig. 3, according to the first left part,
SSAPSO is in the first rank followed by SSA, PSO, GA, and SSAPSO has the least time, then PSO, followed by a BAT,
BAT in terms of the F-measure, and according to the accu- followed by SSA, and GA has the last rank, where it takes
racy, SSAPSO then GA, SSA, BAT, PSO. For the Clean1 the most time. According to the second part of the root

Fig. 3  The time, the root mean

square error(RMSE), and the
selection ratio for all algorithms
over all datasets

13
 R. A. Ibrahim et al.

Fig. 4  The average of the accu-

racy and the F-measures for all
algorithms over all datasets

mean square error, SSAPSO has the least RMSE, followed Table 8  The results of the Wilcoxon test between the SSAPSO and
by PSO, then GA, then SSA and BAT have the most ratio the other algorithms
of error. And according to the third part, the selection ratio, Dataset Measure SSA GA PSO BAT
SSAPSO has the least ratio of the selection of the features,
followed by SSA, followed by GA, then PSO and BAT have WBDC Time 0.147 0.003 0.599 0.147
the greatest ratio of the selection of the features. Thus, the RMSE 0.097 0.000 0.000 0.000
Accuracy 0.097 0.000 0.000 0.000
proposed algorithm takes the least computational time with
Ionosphere Time 0.437 0.365 0.052 0.968
a small number of selected features to represent the whole
RMSE 0.000 0.001 0.002 0.000
dataset regarding the selection ratio.
Accuracy 0.000 0.000 0.014 0.000
In addition, from the left part of the Fig. 4, SSAPSO has
Hepatitis Time 0.179 0.000 0.368 0.368
the greatest value of the average of the accuracy, followed
RMSE 0.812 0.338 0.016 0.102
by SSA followed by PSO, then GA, and BAT has the last
Accuracy 0.090 0.000 0.026 0.000
rank with small accuracy. The same order of the average of
Heart Time 0.201 0.227 0.002 0.000
the F-measure, where SSAPSO has the greatest average of
RMSE 0.047 0.016 1.000 0.000
the F-measure value over all the datasets. Therefore, it can
Accuracy 0.047 0.000 0.002 0.000
conclude that the proposed SSAPSO algorithm is a power-
Breastcancer Time 0.197 0.000 0.000 0.000
ful algorithm in the FS problems since it has the advantages
RMSE 0.002 0.000 0.000 0.0007
of the two algorithms, SSA and PSO algorithms together.
Accuracy 0.002 0.027 0.062 0.000
However, the main limitations of the proposed algorithm
Sonar Time 0.000 0.002 0.000 0.000
are that the selected features are different with each run so,
RMSE 0.000 0.000 0.000 0.018
we take the intersection between the selected features along
Accuracy 0.017 0.334 0.965 0.085
with all runs.
Lymphography Time 0.263 0.000 0.681 0.095
RMSE 0.550 0.000 0.681 0.000
4.5 Statistical analysis
Accuracy 0.007 0.000 0.000 0.000
Clean 1 Time 0.000 0.000 0.000 0.000
In this section, a further statistical analysis is added by
RMSE 0.000 0.000 0.930 0.000
using a non-parametric statistical test called the Wilcoxon’s
Accuracy 0.000 0.000 0.930 0.000
rank sum test. This test is used to determine if the differ-
Clean 2 Time 0.000 0.000 0.000 0.001
ence between the algorithms is significant or not significant
RMSE 0.020 0.413 0.120 0.001
according to the significance level (here .05%). In general,
Accuracy 0.028 0.567 0.341 0.002
there are two hypothesizes (1) null hypothesis that assumes
Waveform Time 0.730 0.000 0.000 0.000
that there is no significant difference between the algo-
RMSE 0.033 0.103 0.016 0.031
rithms; meanwhile, (2) alternative hypothesis that assumes
Accuracy 0.050 0.093 0.142 0.338
that there is a significant difference. The null hypothesis is

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

accepted if the p-value is greater than the significance level; and BAT. According to the performance measures like the
otherwise, the alternative hypothesis is accepted. The results accuracy, the selected feature ratio, the CPU time and other
of Wilcoxon’s rank sum test are listed in Table 8 regarding measures to evaluate the fitness function, the classification
the CPU time(s), the RMSE, and the Accuracy. From this performance of SSAPSO is better than the accuracy of the
table, it can be seen that there exists no significant difference other algorithms.
between the proposed SSAPSO algorithm and the SSA algo- In future works, we will attempt to use the proposed algo-
rithm regarding the CPU time(s) except Clean 1 and Clean rithm in other applications related to different fields such
2 datasets. Also, regarding RMSE, there is a significant dif- as galaxy image classification, QSAR model, renewable
ference between them overall the datasets except for WBDC, energy, image processing and signal processing. As well as,
Hepatitis, and Lymphography datasets, as well as, there is improve the proposed algorithm to work as a multi-objective
no significant difference regarding the accuracy only in two algorithm to solve the big data optimization problem, and
datasets that are called WBDC, and Hepatitis. In addition, task scheduling in cloud computing.
by comparing the GA with the proposed SSAPSO algorithm,
it can be observed that there exists a significant difference Acknowledgements This work is supported by the Science and
Technology Program of Shenzhen of China under Grant Nos.
overall the datasets along all the performance measures. JCYJ20170818160208570 and JCYJ20170307160458368.
Except, for the datasets of Ionosphere and Heart (regarding
CPU time(s)), Hepatitis, Clean 2, and waveform (in terms of
RMSE), and sonar, Clean 2; and waveform (regarding accu- References
racy). Moreover, there is a significant difference between the
proposed algorithm and PSO, regarding the CPU time(s) Al-Ayyoub M, Jararweh Y, Rababah A, Aldwairi M (2017) Feature
along the WBDC, Ionosphere, Hepatitis, and Lymphography extraction and selection for arabic tweets authorship authentica-
datasets. While, based on the RMSE along Lymphography, tion. J Ambient Intell Humaniz Comput 8(3):383–393
Anderson PA, Bone Q (1980) Communication between individu-
Clean 1, and Clean 2 datasets, also, regarding Accuracy als in salp chains II. Physiology. Proc R Soc Lond B Biol Sci
along Breastcancer, sonar, Clean 1, and Clean 2 and wave- 210(1181):559–574
form datasets. Finally, by comparing the proposed SSAPSO Arigbabu OA, Mahmood S, Ahmad SMS, Arigbabu AA (2016)
with the BAT algorithm, it can be concluded that there exists Smile detection using hybrid face representation. J Ambient
Intell Humaniz Comput 7(3):415–426
a significant difference overall the datasets and along each Awada W, Khoshgoftaar TM, Dittman D, Wald R, Napolitano A
measure. Except, regarding CPU time(s), there is no signifi- (2012) A review of the stability of feature selection techniques
cant difference among the datasets of WBDC, Ionosphere, for bioinformatics data. In: 2012 IEEE 13th international con-
Hepatitis, and Lymphography datasets, and according to the ference on information reuse and integration (IRI). IEEE, pp
356–363
accuracy of and sonar, waveform datasets. Chang PC, Lin JJ, Liu CH (2012) An attribute weight assignment
and particle swarm optimization algorithm for medical database
classifications. Comput Methods Prog Biomed 107(3):382–392
Chen LH, Yang B, jing Wang S, Wang G, zhong Li H, bin Liu W
5 Conclusions and future work (2014) Towards an optimal support vector machine classifier
using a parallel particle swarm optimization strategy. Appl Math
This paper introduces an alternative hybrid swarm (called Comput 239:180–197
SSAPSO) that combines the salp swarm algorithm(SSA) and Chikh R, Chikhi S (2017) Clustered negative selection algorithm and
fruit fly optimization for email spam detection. J Ambient Intell
the particle swarm optimization (PSO) algorithm, and it is Humaniz Comput. https​://[Link]/10.1007/s1265​2-017-0621-2
used to enhance the ability of the SSA algorithm for the Chuang LY, Yang CH, Yang CH (2009) Tabu search and binary par-
exploration and the exploitation. The proposed algorithm ticle swarm optimization for feature selection using microarray
utilizes the characteristics of the PSO strategy to improve the data. J Comput Biol 16(12):1689–1703
Cuevas E, Cienfuegos M (2014) A new algorithm inspired in the
quality of the SSA in searching for the solutions. Therefore, behavior of the social-spider for constrained optimization.
the convergence rate is increased. The performance of the Expert Syst Appl 41(2):412–425
SSAPSO algorithm is evaluated under two experiments, in Dash M, Liu H (2003) Consistency-based search in feature selection.
the first experiment, it is tested over a set of 15 benchmark Artif Intell 151(1–2):155–176
Eberhart R, Kennedy J (1995) A new optimizer using particle swarm
functions and compares its performance with the standard theory. In: Micro machine and human science, 1995. MHS’95.
SSA through several measures. Proceedings of the Sixth International Symposium on, IEEE,
To further evaluate the performance of the SSAPSO, in pp 39–43
the second experiment, we applied it to select the optimal El Aziz MA, Ewees AA, Hassanien AE (2016) Hybrid swarms opti-
mization based image segmentation. In: Bhattacharyya S, Dutta
subset of features from ten UCI datasets to increase the accu- P, De S, Klepac G (eds) Hybrid soft computing for image seg-
racy of the classification. In which, the SSAPSO is com- mentation. Springer, Berlin, pp 1–21
pared with other four algorithms namely, SSA, GA, PSO, El Aziz MA, Hemdan AM, Ewees AA, Elhoseny M, Shehab A, Has-
sanien AE, Xiong S (2017) Prediction of biochar yield using

13
 R. A. Ibrahim et al.

adaptive neuro-fuzzy inference system with particle swarm opti- and Communications, 2006. ADCOM 2006. International Con-
mization. In: PowerAfrica, 2017 IEEE PES, IEEE, pp 115–120 ference on, IEEE, pp 298–303
El Aziz MA, Ewees AA, Hassanien AE (2018a) Multi-objective Kira K, Rendell LA (1992) The feature selection problem: traditional
whale optimization algorithm for content-based image retrieval. methods and a new algorithm. In: Proc AAAI 1992, San Jose,
Multimed Tools Appl 77:26135–26172 CA, pp 129–134
El Aziz MA, Ewees AA, Hassanien AE, Mudhsh M, Xiong S (2018b) Kohane IS, Butte AJ, Kho A (2002) Microarrays for an integrative
Multi-objective whale optimization algorithm for multilevel genomics. MIT press, Cambridge
thresholding segmentation. In: Hassanien A, Oliva D (eds) Kohavi R (1994) Feature subset selection using the wrapper method,
Advances in soft computing and machine learning in image overfitting and dynamic search space topology. In: Proc AAAI
processing. Springer, Berlin, pp 23–39 Fall Symposium on Relevance, pp 109–113
Elaziz MEA, Ewees AA, Oliva D, Duan P, Xiong S (2017) A hybrid Kung SY, Luo Y, Mak MW (2010) Feature selection for genomic sig-
method of sine cosine algorithm and differential evolution for nal processing: unsupervised, supervised, and self-supervised
feature selection. In: Liu D, Xie S, Li Y, Zhao D, El-Alfy ES scenarios. J Signal Process Syst 61(1):3–20
(eds) International conference on neural information processing. Lai C, Reinders MJ, Wessels L (2006) Random subspace method
Springer, Berlin, pp 145–155 for multivariate feature selection. Pattern recognition letters
Ewees AA, El Aziz MA, Elhoseny M (2017a) Social-spider optimi- 27(10):1067–1076
zation algorithm for improving anfis to predict biochar yield. In: Li X, Wang G (2015) Optimal band selection for hyperspectral data
2017 8th International Conference on Computing, Communica- with improved differential evolution. J Ambient Intell Humaniz
tion and Networking Technologies (ICCCNT), IEEE, pp 1–6 Comput 6(5):675–688
Ewees AA, El Aziz MA, Hassanien AE (2017b) Chaotic multi-verse Li J, Wong L (2002) Identifying good diagnostic genes or genes
optimizer-based feature selection. Neural Comput Appl. https​:// groups from gene expression data by using the concept of
[Link]/10.1007/s0052​1-017-3131-4 emerging patterns. Bioinformatics 18:725–734
Ewees AA, Elaziz MA, Houssein EH (2018) Improved grasshopper Liu Y, Wang G, Chen H, Dong H, Zhu X, Wang S (2011) An
optimization algorithm using opposition-based learning. Expert improved particle swarm optimization for feature selection. J
Syst Appl 112:156–172 Bionic Eng 8(2):191–200
Gasca SJARE (2006) Eliminating redundancy and irrelevance using Madin LP (1990) Aspects of jet propulsion in salps. Can J Zool
a new mlp-based feature selection method. Pattern Recognit 68(4):765–777
39(2):313–315 Menghour K, Souici-Meslati L (2016) Hybrid aco-pso based
Goldberg DE, Holland JH (1988) Genetic algorithms and machine approaches for feature selection. Int J Intell Eng Syst 9(3):65–79
learning. Machine Learn 3(2):95–99 Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer. Adv
Guyon I, Elisseeff A (2003) An introduction to variable and feature Eng Softw 69:46–61
selection. J Mach Learn Res 3:1157–1182 Mirjalili S, Gandomi AH, Mirjalili SZ, Saremi S, Faris H, Mirjalili
Guyon I, Weston J, Barnhill S, Vapnik V (2002) Gene selection SM (2017) Salp swarm algorithm: a bio-inspired optimizer for
for cancer classification using support vector machines. Mach engineering design problems. Adv Eng Softw 114:163–191
Learn 46(1–3):389–422 Modrzejewski M (1993) Feature selection using rough sets theory.
Hafez AI, Hassanien AE, Zawbaa HM, Emary E (2015) Hybrid mon- In: Proceedings of the European Conference on Machine Learn-
key algorithm with krill herd algorithm optimization for feature ing, Vienna, Austria, pp 213–226
selection. In: 2015 11th International computer engineering Moradi P, Rostami M (2015) Integration of graph clustering with
conference (ICENCO). IEEE, pp 273–277 ant colony optimization for feature selection. Knowl Based Syst
Han J, Pei J, Kamber M (2011) Data mining: concepts and tech- 84:144–161
niques. Elsevier, New York Neumann J, Schnörr C, Steidl G (2005) Combined svm-based feature
Henschke N, Everett JD, Richardson AJ, Suthers IM (2016) selection and classification. Mach Learn 61(1–3):129–150
Rethinking the role of salps in the ocean. Trends Ecol Evol Niknam T, Amiri B (2010) An efficient hybrid approach based on
31(9):720–733 pso, aco and k-means for cluster analysis. Appl Soft Comput
Ibrahim RA, Oliva D, Ewees AA, Lu S (2017) Feature selection 10(1):183–197
based on improved runner-root algorithm using chaotic singer Noman S, Shamsuddin SM, Hassanien AE (2009) Hybrid learning
map and opposition-based learning. In: Liu D, Xie S, Li Y, Zhao enhancement of rbf network with particle swarm optimization.
D, El-Alfy ES (eds) International conference on neural informa- In: Hassanien AE, Abraham A, Vasilakos AV, Pedrycz W (eds)
tion processing. Springer, Berlin, pp 156–166 Foundations of computational, intelligence, vol 1. Springer,
Ibrahim RA, Elaziz MA, Ewees AA, Selim IM, Lu S (2018) Galaxy Berlin, pp 381–397
images classification using hybrid brain storm optimization Prabukumar M, Agilandeeswari L, Ganesan K (2017) An intelligent
with moth flame optimization. J Astron Telesc Instrum Syst lung cancer diagnosis system using cuckoo search optimization
4(3):038001 and support vector machine classifier. J Ambient Intell Humaniz
Inbarani HH, Azar AT, Jothi G (2014) Supervised hybrid feature Comput. https​://[Link]/10.1007/s1265​2-017-0655-5
selection based on pso and rough sets for medical diagnosis. Raymer ML, Punch WF, Goodman ED, Kuhn LA, Jain AK (2000)
Comput Methods Prog Biomed 113(1):175–185 Dimensionality reduction using genetic algorithms. IEEE trans-
Jensen R, Goodarzi M, Freitas MP (2009) Feature selection and lin- actions on evolutionary computation 4(2):164–171
ear/nonlinear regression methods for the accurate prediction of Rodrigues D, Pereira LA, Nakamura RY, Costa KA, Yang XS, Souza
glycogen synthase kinase-3beta inhibitory activities. J Chem AN, Papa JP (2014) A wrapper approach for feature selection
Inf Model 49:824–832 based on bat algorithm and optimum-path forest. Expert Syst
Karaboga D, Basturk B (2007) A powerful and efficient algorithm Appl 41(5):2250–2258
for numerical function optimization: artificial beecolony (abc) Saravanan RA, Rajesh Babu M (2017) Enhanced text mining
algorithm. J Global Optim 39(3):459–471 approach based on ontology for clustering research project
Karnan M, Thangavel K, Sivakuar R, Geetha K (2006) Ant colony selection. J Ambient Intell Humaniz Comput. https​: //doi.
optimization for feature selection and classification of microc- org/10.1007/s1265​2-017-0637-7
alcifications in digital mammograms. In: Advanced Computing

13
Improved salp swarm algorithm based on particle swarm optimization for feature selection

Sutherland KR, Weihs D (2017) Hydrodynamic advantages of swim- Yamuna G, Thamaraichelvi B (2016) Hybrid firefly swarm intelli-
ming by salp chains. J R Soc Interface 14(133):20170,298 gence based feature selection for medical data classification and
Tanaka K, Kurita T, Kawabe T (2007) Selection of import vectors via segmentation in svd–nsct domain. Int J Adv Res 4(9):744–760
binary particle swarm optimization and cross-validation for ker- Yang Y, Slattery S, Ghani R (2002) A study of approaches to hyper-
nel logistic regression. In: Proceedings of the International Joint text categorization. J Intell Inform Syst 18(2):219–241
Conference on Neural Networks (IJCNN 07),IEEE, Orlando, Yao YY (2003) Information-theoretic measures for knowledge
Fla, USA, pp 1037–1042 discovery and data mining. In: Karmeshu (ed) Entropy meas-
Thangavel K, Velayutham C (2011) Mammogram image analysis: ures, maximum entropy principle and emerging applications.
bioinspired computational approach. In: Proceedings of the Springer, Berlin, Heidelberg, pp 115–136
International Conference on Soft Computing for Problem Solv- Yeh WC, Yang YT, Lai CM (2016) A hybrid simplified swarm opti-
ing, pp 20–22 mization method for imbalanced data feature selection. Aust
Unler A, Murat A (2010) A discrete particle swarm optimization Acad Bus Econ Rev 2(3):263–275
method for feature selection in binary classification problems. Zhang H, Sun G (2002) Feature selection using tabu search method.
Eur J Oper Res 206(3):528–539 Pattern Recognit 35(3):701–711
Wang Y, Cen Y, Zhao R, Zhang L, Kan S, Hu S (2018) Compressed Zhong DJN (2001) Using rough sets with heuristics for feature selec-
sensing based feature fusion for image retrieval. J Ambient Intell tion. J Intell Inform Syst 16:199–214
Humaniz Comput. https​://[Link]/10.1007/s1265​2-018-0895-z
Xie ZX, Hu QH, Yu DR (2006) Improved feature selection algorithm Publisher’s Note Springer Nature remains neutral with regard to
based on svm and correlation. In: International Symposium on jurisdictional claims in published maps and institutional affiliations.
Neural Networks. Springer, Berlin, pp 1373–1380

13