Co-Operative Binary Bat Optimizer with Rough Set Reducts for Text Feature Selection

Abstract

The process of eliminating irrelevant, redundant and noisy features while trying to maintain less information loss is known as a feature selection problem. Given the vast amount of the textual data generated and shared on the internet such as news reports, articles, tweets and product reviews, the need for an effective text-feature selection method becomes increasingly important. Recently, stochastic optimization algorithms have been adopted to tackle this problem. However, the efficiency of these methods is decreased when tackling high-dimensional problems. This decrease could be attributed to premature convergence where the population diversity is not well maintained. As an innovative attempt, a cooperative Binary Bat Algorithm (BBA_CO) is proposed in this work to select the optimal text feature subset for classification purposes. The proposed BBA_CO uses a new mechanism to control the population’s diversity during the optimization process and to improve the performance of BBA-based text-feature selection method. This is achieved by dividing the dimension of the problem into several parts and optimizing each of them in a separate sub-population. To evaluate the generality and capability of the proposed method, three classifiers and two standard benchmark datasets in English, two in Malay and one in Arabic were used. The results show that the proposed method steadily improves the classification performance in comparison with other well-known feature selection methods. The improvement is obtained for all of the English, Malay and Arabic datasets which indicates the generality of the proposed method in terms of the dataset language.

1. Introduction

Text classification is the process of automatic grouping of documents into some predefined categories. The idea of text classification is to assign one document to one class (i.e., category), based on its contents. It can provide conceptual views of document collection and has essential applications in the real world. For example, news stories are typically organized by subject categories (topics) or geographical codes; academic papers are often classified by technical domains and sub-domains; even patient reports in health-care organizations are often indexed from multiple aspects, using taxonomies of disease categories, types of surgical procedures, insurance reimbursement codes and so on.

Text Feature Selection (TFS) is an important part of text classification, and much research has been completed on various feature selection methods. A document usually contains hundreds or thousands of distinct words regarded as features. However, many of them may be noisy, less informative or redundant with respect to the class label. This may mislead the classifiers and degrade their performance in general [1,2]. Feature selection (FS) can be thought of as selecting the best words of a document that can help classify that document. Feature selection has been an active research area in pattern recognition, machine learning, statistics and data mining communities. The main idea of feature selection is to choose a subset of the original features by eliminating redundant ones and those with little or no predictive information. Feature selection is an essential process, as it can make or break a classification engine [3]. Feature selection is considered an optimization problem [4,5,6] where the aim is to select the most representative features that give the highest prediction performance. The idea of TFS, in simple words, is to determine the importance of words using a defined measure that can keep informative words, and remove non-informative words, which can then help the text classification engine.

During the past few decades, many feature selection methods have been proposed. On one hand, some of those methods work by ranking features and filtering out the low-ranked ones. Although those methods are fast and independent from any classification algorithm, they ignore the dependencies between features which affect the quality of the selected feature set [7]. On the other hand, population-based meta-heuristic methods such as genetic algorithm (GA), ant colony optimization (ACO), particle swarm optimization (PSO) and Bat algorithm (BA) have attracted a lot of attention [5,7,8,9,10,11,12,13,14,15,16,17,18]. These methods try to gather better solutions by using knowledge from previous steps. Therefore, the focus on search strategies has shifted to meta-heuristic algorithms, which are well suited for searching among a large number of possibilities for solutions. Most of these methods utilizes a classification method to evaluate the feature set, resulting in higher classification accuracy. However, the main drawback of these methods is that they are dependent on the utilized classification algorithm, and this makes the resulted feature sets biased to the choice of classifier [19].

The Bat algorithm (BA) is a meta-heuristic method proposed by [20] and based on the fascinating capability of micro-bats to find their prey and discriminate different types of insects even in complete darkness. The algorithm is formulated to imitate the ability of bats in finding their prey. The main advantage of the BA is that it combines the benefits of population-based and single-based algorithms to improve the quality of convergence [21]. BA and its variants have been successfully applied to solve many problems such as optimization, classification, feature selection, image processing and scheduling [8,21,22,23,24,25,26]. For more details about the Bat algorithm and the binary version of it, the reader may refer to [20,27], respectively.

As mentioned above, BA was successfully applied to many application domains including FS. However, one of the limitations with many meta-heuristic algorithms, including BA, is their deficient performance with high-dimensional problems. This problem is most likely to appear as the search space is not effectively explored due to losing population diversity during the search process [28]. Many methods were proposed in the literature to control population diversity including cooperative algorithms [29,30]. However, for high-dimensional problems, co-evolutionary algorithms are preferred as they can divide the dimension of the solution into multiple parts, and optimize each part separately [31,32]. Moreover, as the text data are represented as a sequence of terms where each term is considered as one feature, this aggravates the problem of high dimensionality. The coevolutionary strategy was successfully employed in several evolutionary computations, such as job-shop scheduling [33], path-planning problem [34], supply chain-gap analysis [35], flow-shop scheduling problem [36], large-scale optimization [37], hierarchized Steiner tree problems [38] and sensor ontology meta-matching [39]. However, the majority of these applications are continuous problems. Applying this technique to a discrete problem such as text feature selection is still challenging and needs to be further studied.

In this paper, a cooperative coevolutionary BBA is proposed and evaluated as a TFS method, that provides the following contributions:

• Controlling the population diversity during the search process using the multi-population BBA;
• Handling the high dimensionality of the feature space by using the divide and conquer strategy;
• Initializing a diverse population using the modified Latin Hypercube Sampling (LHS) initialization method;
• Better evaluation of the solutions using the adapted Rough Set (RS)-based fitness function that is independent of any classification method.

The rest of this paper is organized as follows: Section 2 provides a summary of the related work. Then, the details of the proposed algorithm are given in Section 3, followed by the experimental setup in Section 4. After that, the discussion and analysis of the experimental results are shown in Section 5. Finally, the work is concluded in Section 6 of this paper.

2. Related Work

The simplest definition of a coevolutionary algorithm is that it is an evolutionary algorithm (or a collection of evolutionary algorithms) in which the fitness of an individual depends on the relationship between that individual and other individuals [40]. Such a definition immediately imbues these algorithms with a variety of views differing from those of more traditional evolutionary algorithms. Therefore, the interaction between individuals of different populations is the key to the success of coevolutionary techniques.

In the literature, coevolution is often divided into two classes: cooperative and competitive, regarding the type of interaction employed. In cooperative coevolution, each population evolves individuals representing a component of the final solution. Thus, a full candidate solution is obtained by joining an individual chosen from each population. In this way, increases in a collaborative fitness value are shared among individuals of all the populations of the algorithm. In competitive coevolution, the individuals of each population compete with each other. This competition is usually represented by a decrease in the fitness value of an individual when the fitness value of its antagonist increases [41].

Additionally, coevolution is a research field that has recently started to grow. Some research efforts have been applied to tackle the question about how to select the members of each population that will be used to evaluate the fitness function. One way is to evaluate an individual against every single collaborator in the other population. Although it could be a better way to select the collaborators, it would consume a very high number of evaluations in the computation of the fitness function. To reduce this number, there are other options, such as the use of just a random individual or the use of the best individual from the previous generation [42].

The coevolutionary technique was successfully utilized in literature with different domains. In an early work, the authors of [43] presented the cooperative particle swarm optimizer and applied their method to several benchmark optimization problems. The authors of [44] proposed an approach based on coevolutionary particle swarm optimization to solve constrained optimization problems formulated as min–max problems. Another study [45] proposed a cooperative coevolution framework in order to optimize large scale non-separable problems. The authors of [46] adapted a competitive and cooperative coevolutionary approach for a multi-objective particle swarm optimization algorithm design, which appeared to solve complex optimization problems by explicitly modelling the coevolution of competing and cooperating species. In another work, the authors of [47] proposed a cooperative coevolving particle swarm optimization algorithm in an attempt to address the issue of scaling-up particle swarm optimization algorithms in solving large-scale optimization problems.

Later, the authors of [48] proposed a direction vector-based coevolutionary multi-objective optimization algorithm, that introduced the decomposition idea from multi-objective evolutionary algorithms to coevolutionary algorithms. The authors of [49] proposed an adaptive coevolutionary algorithm based on genotypic diversity measure. In another study [50] a coevolutionary improved multi-ant colony optimization algorithm was proposed for ship multi and branch-pipe route design. The author of [51] proposed a cooperative coevolutionary artificial bee colony algorithm that has two sub-swarms, with each addressing a sub-problem. The sub-problems were a charge scheduling problem in a hybrid flow-shop, and a cast scheduling problem in parallel machines.

Later, the authors of [52] proposed a multi-objective cooperative coevolutionary algorithm to optimize the reconstruction term, the sparsity term and the total variation regularization term, simultaneously, for Hyperspectral Sparse Unmixing. In [53] the authors proposed a parallel multi-objective cooperative coevolutionary variant of the Speed-constrained Multi-objective Particle Swarm Optimization algorithm. In [54], the authors proposed a two-layer distributed cooperative coevolution architecture with adaptive computing resource allocation for large-scale optimization. In another study, [55], the authors proposed an approach utilizing a Cooperative Co-evolutionary Differential Evolution algorithm to optimize high-dimensional ANNs.

In a recent study [56], the authors proposed a hybrid cooperative coevolution algorithm for the minimization of fuzzy makespan. In [57], the authors developed a cooperative coevolution algorithm for seru production with minimizing makespan by solving the seru formation and seru scheduling simultaneously. In another study [58] the authors proposed a multi-population coevolution-based multi-objective particle swarm optimization algorithm to realize the rapid search for the globally optimal solution to solve the problem of Weapon–Target Assignment. In addition to these studies, the authors of [59] proposed a cooperative coevolution hyper-heuristic framework to solve workflow scheduling problem with an objective of minimizing the completed time of workflow.

For feature selection problems, a few studies in the literature have utilized the cooperative coevolutionary algorithm. In two early works, refs. [60,61], the authors performed instance and feature selection by creating three populations in different sizes. The first population performed feature selection, while the second population performed instance selection and the third population was for both feature and instance selection. The authors of [62] presented a hybrid learning algorithm based on a cooperative coevolutionary algorithm (Co-CEA) with dual populations for designing the radial basis-function neural network (RBFNN) models with an explicit feature selection. In this algorithm, the first sub-population used binary encoding masks for feature selection, and the second sub-population tended to yield the optimal RBFNN structure.

Another study presented a cooperative coevolution framework to render the feature selection process embedded into the classification model construction within the genetic-based machine learning paradigm [42]. Their approach had two coevolving populations cooperate with each other regarding the fitness evaluation. The first population corresponded to the selected feature subsets and the second population was for rule sets of classifier. Later, the authors of [63] proposed an attribute equilibrium dominance reduction accelerator (DCCAEDR) based on the distributed coevolutionary cloud model. The framework of N-populations distributed coevolutionary MapReduce model is designed to divide the entire population into N sub-populations, sharing the rewards of different sub-populations’ solutions under a MapReduce cloud mechanism. After that, a CCFS algorithm was proposed that divided vertically (on features) the dataset by random manner and utilized the fundamental concepts of cooperation coevolution in order to search the solution space via Binary Gravitational Search Algorithm (BGSA) [28]. Another study utilized a genetic algorithm for the coevolution of meteorological data for attribute reduction [64]. In this work, the evolutionary population was divided into two sub-populations; one for elite individuals to assist crossover operations to increase the convergence speed of the algorithm, and the other for balancing the population diversity in the evolutionary process by introducing a random population.

It is noticed that in most of the mentioned studies [42,60,61,62,64], the authors have attempted to solve the feature selection problem as a multi-objective problem by creating two or more populations, where each of them optimizes one objective. However, those methods are not applicable for single objective problems and they do not solve the high dimensionality of the feature space. In the work of Ding et al. [63], the focus was to distribute the optimization process on multiple machines in order to reduce the computational time. However, the requirement of their model, such as the hardware (e.g., multiple PC machines), the mechanism of distribution of the dataset, the means of communication between different machines, and the way of forming the complete solution, was not always available. In the work of the authors of [28] the dimension of the full solution was divided into smaller subsets where each of them is optimized in a separate population. Although their method was effective with high dimensional feature selection problem, there were multiple aspects that needed further improvement. For example, the method might have a better parameter tuning in order to improve its performance. In addition, the solutions in the different sub-populations need to be combined with each other in each generation in order to be evaluated, which is computationally expensive and reduces the chance of each solution to be optimized separately from the other sub-populations.

3. The Proposed Algorithm

In this work, the cooperative and coevolution mechanisms are utilized with the binary bat algorithm. A combination of these two approaches (which we call BBA_CO) is proposed for the text feature selection problem. Figure 1 shows the main stages of the proposed method that are explained in the following subsections.

3.1. Initialization Stage

This stage contains three steps, i.e., (i) BBA_CO parameter initialization that is based on the parameter tuning of the Taguchi method, (ii) sub-population initialization using a modified LHS method and (iii) solutions’ evaluation to choose the best candidate solution in each sub-population based on the dependency measurement using rough set theory (RST) [65]. Algorithm 1 shows the pseudocode for the initialization stage.

Algorithm 1. Pseudocode of the initialization stage.

Initialization

//Step 1: BBACO parameter initialization Initialize SubPop-no, SubPop-size, Evaluate-fullSol-rate //Step 2: Sub-population initialization Size-PartialSolution = F/SubPop-no, where F total represents number of features Remainder = F%SubPop-no Foreach Sub-Population (i), where i represent the index from 0 to SubPop-no  Ifi <= Reminder   Size-PartialSolution(i) = Size-PartialSolution + 1  Else   Size-PartialSolution(i) = Size-PartialSolution   For1 to SubPop-size   Generate initial solution using modified LHS method   Initialize loudness (A), pulse rate (r), minimum frequency (Fmin), velocity (v),    maximum frequency (Fmax); //Step 3: Solution evaluation Evaluate each solution in each sub-population using rough-set based objective function Assign the best solution into $x_{best}^i$ Combine all $x_{best}^i$ into $x_{best}$ Save $x_{best}$ in memory

In Step 1, SubPop-no refers to the number of sub-populations, SubPop-size refers to the number of candidate solutions in each sub-population, and Evaluate-fullSol-rate refers to the number of generations reproduced before evaluating the full dimension solutions (referred to as FullSolution), are initialized. In this work, a Taguchi method [66] was used to identify the best values of the parameters for the BBA_CO algorithm. Three levels were considered for each factor as shown in

Table 1. The BBA_CO algorithm parameter levels for the Taguchi method.

Parameter	Definition	Level
		1	2	3
SubPop-size	The sub-population size	50	100	150
SubPop-no	The number of sub-populations	10	15	20
Evaluate-fulSol-rate	The number of reproduced generations before evaluating the full dimension solutions (referred to as FullSolution)	10	20	30

. The BBA_CO algorithm runs three times for each factor at each level, and the average Signal-to-Noise (SN) ratio plot for each level of the factors is shown in Figure 2. The level with the maximum SN ratio is the optimum parameter determined by the Taguchi method. According to Figure 2, the optimum value for SubPop-size is set to 100, the Evaluate-fullSol-rate is set to 10, and the SubPop-no is set to 15 as shown in Table 1.

In Step 2, the candidate solutions of each sub-population are initialized using a modified LHS method. Note that the candidate solution in each sub-population that contains subset of features is referred to as PartialSolution, while FullSolution refers to a candidate solution with a full dimension where its length is equal to the total number of features. The size of PartialSolution (Size-PartialSolution) is determined based on the number of features and the number of sub-populations as in the following equations:

Size-PartialSolution = F/SubPop-no

Remainder = F % SubPop-no

where F is the total number of features, and Remainder is the number of sub-populations that will be assigned extra one feature. For example, if F is 20, and SubPop-no is 3, thus the Size-PartialSolution and Remainder are:

Size-PartialSolution = 20/3 = 6

Remainder = 20%3 = 2

Based on the value of the Remainder, the size of two out of three sub-populations will contain one extra feature. Thus, seven features are yielded for sub-population #1 to #2 (i.e., Size-PartialSolution +1 = 6 + 1 = 7). The remaining one sub-population remains with six features. Figure 4 shows how the full dimension of features, FullSolution (i.e., 20 features) is divided and assigned into three sub-populations where the first two sub-populations consist seven features, and the third sub-population consists of six features. Note that the letters in Figure 3 represent the features where the arrows represent the movement of features to the sub-populations and # is number (#2 equals to number 2).

In Step 3, each candidate solution in all sub-populations is evaluated based on the adapted dependency degree measure using rough set theory. For example, in Figure 4 the selected features are a and d. Thus, the quality of the PartialSolution is calculated based on the dependency degree between feature a and d using the adapted rough set theory (RST). In the adapted RST, the features are represented by their presence or absence in the document. In this way, the candidate solutions could be compared with the instances (i.e., documents) to define the lower and upper approximations. However, it was found that there is no instance that can have the same pattern of the solution in order to be added to a lower approximation due to the high dimensionality of the feature space. To handle this limitation, the similarity between each instance and the candidate solution is calculated using cosine similarity measure as in the following equation. The similarity threshold (δ = 0.70) is defined based on the preliminary experiments:

$$cos\left(d_i,cs\right)=\frac{{\sum }_{k=1}^n{d}_i^k·c{s}^k}{\sqrt{{\sum }_{k=1}^n{\left(d_i^k\right)}^2}·\sqrt{{\sum }_{k=1}^n{\left(c{s}^k\right)}^2}}$$

where d_i is the document number i; cs is the candidate solution, $d_i^k.c{s}^k$ is the dot product between d_i and cs; k is the index of the term and n is the number of the selected terms in the candidate solution. After calculating cosine similarity, if its value is equal or greater than δ the document is added to the lower approximation, otherwise, it is not added.

The best PartialSolution from each sub-population is assigned to $x_{best}^i$, and is concatenated as follows:

$$x_{best}=concatenate\left(x_{best}^1 , x_{best}^2 , x_{best}^3 , \dots , x_{best}^{subPop-no}\right)$$

Modified LHS: Initial Population Generation

Latin Hypercube Sampling (LHS) is a statistical method developed by the author of [67] and used for sampling by ensuring that all portions of the continuous variable were sampled. LHS was used as an initialization method by the authors of [68], where the solutions were represented using real values as the problem was continuous. For initializing the candidate solutions of BBACO, the LHS method was modified to be applicable to the problems with binary representation. The modified LHS works as follows:

• Divide the length of the solution into equal segments, where the length of the solution is equal to the number of features. The following equation is used to determine the number of segments in each solution:
  $$sn=\frac{F}{n}\mathit{random}(1,m)$$

  $$m=\left\{\begin{array}{l}\frac{F}{n} \mathit{if}\frac{F}{n}\le \frac{n}{2}\\ \frac{n}{2} \mathit{if}\frac{F}{n}>\frac{n}{2}\end{array}\right.$$

  where sn refers to the number of segments; F is the number of features and n is the number of solutions in the population. $\frac{F}{n}$ provides the number of segments that guarantees using each feature only one time in one solution. The parameter m is the upper band of the random number, and it ensures that the number of the selected features does not exceed half of the features. It should be noted that m is defined one time at the beginning of the initialization process. The reason behind using two different ways to calculate m (depending on the size of features and population) is to make the method more suitable with datasets of a different size;
• Calculate the length of the segments (sl) for each solution as follows:
  $$sl=\frac{F}{sn}$$

Then, one feature is selected randomly from each segment. The steps of the modified LHS are shown in Algorithm 2.

Algorithm 2. The modified LHS initialization method steps.

Modified LHS initialization method

Calculate m, where m is the maximum number of the features that can be selected   For 1 to number of solutions Calculate the number of segments (sn) Calculate the length of segment (sl) Randomly choose one feature in each segment. Check if this is the final solution:    If yes: go to solutions’ evaluation.    Otherwise: go to 2

3.2. Improvement Stage

This stage contains two steps, i.e., (i) local search (random walk) and (ii) global search (random fly). The local search is applied on each $x_{best}^i$ (best PartialSolution of each sub-population i), while the global search is applied on all PartialSolutions in all sub-populations with the aim of reproduction. In step i, a local PartialSolution is generated based on the best PartialSolution in each sub-population if the condition of local search (i.e., $r_i>rand\left[0,1\right]$, where $r_i$ is the pulse rate of the best PartialSolution in sub-population #i) is met. The pseudocode in Algorithm 3 shows the steps of local search.

Algorithm 3. Pseudocode of the local search.

Local search (random walk)

For each $x_{best}^i$    If (ri > rand [0, 1]) $$\begin{gathered} x_{new}^i=x_{old}^i+ϵ\overline{A_g} \\ S\left(x_{new}^i\right)=\frac{1}{1+e^{-x_{new}}} \\ {x}_{new}^i=\{\begin{array}{c}1 if S\left(x_{new}^i\right)>\sigma \left[0,1\right]\\ 0 otherwise\end{array} \end{gathered}$$

In Algorithm 3, r_i is the pulse rate of the best PartialSolution in sub-population #i, rand, $ϵ$, and σ are random numbers between 0 and 1, $x_{new}^i$ is the generated PartialSolution near $x_{best}^i$ (i.e., 15 local PartialSolutions are generated as the number of sub-populations is 15), $\overline{A_g}$ is the average loudness of all PartialSolutions at the generation g, and $S\left(x_{new}^i\right)$ is the sigmoid function that used to restrict the values of $x_{new}^i$ into 0 or 1.

Step 2 generates new PartialSolutions for the next generation. First, the frequency of each PartialSolution in each sub-population is updated. Then, the PartialSolution’s velocity is updated based on the new value of frequency; the best PartialSolution in the corresponding sub-population, and the position (i.e., the PartialSolution itself). After that, the original velocity and the current position are used to generate a new position (i.e., PartialSolution). These operations are shown in the pseudocode of global search in Algorithm 4.

Algorithm 4. Pseudocode of the global search.

Global search (random fly)

For each PartialSolution(pSol) in each sub-population (i) $$\begin{gathered} \hspace{1em}\hspace{1em}{f}_i^{sol}=f_{min}+\left(f_{max}-f_{min}\right)\beta \left[0,1\right] \\ \hspace{1em}\hspace{1em}{v}_i^g\left(pSol\right)=v_i^{g-1}\left(pSol\right)+\left(x_i^{g-1}\left(pSol\right)-x_{best}^i\right)f_i^{pSol} \\ \hspace{1em}\hspace{1em}{x}_i^g\left(pSol\right)=x_i^{g-1}\left(pSol\right)+v_i^g\left(pSol\right) \\ \hspace{1em}\hspace{1em}S\left(x_i^g\left(pSol\right)\right)=\frac{1}{1+e^{-x_i^g\left(pSol\right)}} \\ \hspace{1em}\hspace{1em}{x}_i^g\left(pSol\right)=\{\begin{array}{c}1 if S\left(x_i^g\left(pSol\right)\right)>\sigma \\ 0 otherwise\end{array} \\ \hspace{1em}\hspace{1em}\mathrm{if}(A_i^{pSol}\langle random\left[0,1\right]\& fit\left(x_i^g\left(pSol\right)\right)\rangle fit \left(x_i^{g-1}\left(spSolol\right)\right) \\ \mathit{Accept}\mathit{the}\mathit{new}\mathit{PartialSolution} \\ \hspace{1em}\hspace{1em}\hspace{1em}{r}_i^{g+1}=r_i^0\left[1-exp\left(-\gamma g\right)\right] \\ \hspace{1em}\hspace{1em}\hspace{1em}{A}_i^{g+1}=\alpha A_i^g \mathit{where} \alpha =\gamma =0.9 \end{gathered}$$

As shown in Algorithm 4, the frequency of each PartialSolution is updated as in the second line in the pseudocode where $f_i^{pSol}$ is the new frequency of the PartialSolution in sub-population i, $f_{min}$ is the minimum frequency, $f_{max}$ is the maximum frequency and β is a random number between 0 and 1. Then, the velocity is updated as in the third line in the pseudocode, where $v_i^g\left(pSol\right)$ is the velocity of the PartialSolution in generation g and sub-population i, $v_i^{g-1}\left(pSol\right)$ is the velocity of the PartialSolution in the previous generation and $x_{best}^i$ is the best PartialSolution in sub-population i. After that, a new position (i.e., PartialSolution) is generated as in the fourth and fifth lines in the pseudocode, based on the PartialSolution in the previous generation and the velocity. It could be noted that for the position, the sigmoid function is used to restrict the new values into 0 or 1. The last part of this step is to update the PartialSolutions in the sub-population. If the condition of accepting PartialSolution is met (i.e., loudness (A_i) is less than random number from 0 to 1, and the new PartialSolution is better than the previous one based on the adapted dependency measure of rough set theory, which was utilized as fitness function), then, the PartialSolution is accepted and the pulse rate and loudness are updated.

3.3. Cooperative Stage: FullSolutions Evaluation

This stage takes place after every 10 generations (i.e., Evaluate-fullSol-rate = 10, as explained in Section 3.1) of the algorithm. The parameter Evaluate-fullSol-rate determines which generations of the PartialSolutions in different sub-populations will be concatenated and evaluated as FullSolutions. The purpose of this stage is to cooperate between all sub-populations. The cooperation is achieved by concatenating the PartialSolution in hand (e.g., a PartialSolution in sub-population #1) with the best PartialSolutions of other sub-populations (e.g., sub-population #2 to #15) in generation g, so that the whole sub-populations cooperate with each other to evaluate the PartialSolutions. If the condition of evaluating FullSolution is met (i.e., generation % Evaluate-fullSol-rate = 0), then the PartialSolutions are evaluated by following the pseudocode in Algorithm 5, as shown in Figure 5.

Algorithm 5. Pseudocode of the cooperative stage.

Cooperative stage

if (generation % 10 == 0)   for each sub-population (i)    for each PartialSolution     concatenate with the best solutions of other sub-populations     evaluate FullSolution     update the fitness of the PartialSolution      update the best PartialSolution$x_{best}^i$

As an example for the cooperative stage, suppose that there are three sub-populations, each with three PartialSolutions. Each PartialSolution in sub-population #1 focuses on the best PartialSolutions in sub-populations #2 and #3. Then, the FullSolution with the full dimension (i.e., length = 20) is evaluated using RST-based fitness function. The quality of the PartialSolution is updated based on the new evaluation. The same steps explained in this subsection, are repeated for the PartialSolutions of the other sub-populations. The last step in this stage is to update the best PartialSolution of each sub-population depending on the new fitness values of the PartialSolutions.

3.4. Selection of FinalSolution

This stage takes place when the stopping criteria (i.e., Iter-no >100), is met. In this stage, the saved FullSolutions after each generation are evaluated and the best one is selected as a final solution. As mentioned in the previous sections, after each generation, the best PartialSolutions of all sub-populations are concatenated and saved in the memory. The importance of this stage is due to the nature of the saved FullSolutions as they consist of several parts from several sub-populations. Although each part of these FullSolutions was evaluated and selected as the best PartialSolution within its sub-population in a certain generation, the FullSolution was not evaluated as one set of features. Therefore, to ensure that the final solution is the best one obtained by the algorithm, the saved FullSolutions are evaluated and the best FullSolution is selected as the final solution. Algorithm 6 shows the steps of selecting the final solution.

Algorithm 6. Steps of selection of the final solution.

Selection of final solution

If (iter-no >100)    For i = 1 to 100      Evaluate the saved FullSolution #i    Select the best FullSolution as the final solution

4. Experimental Setup

This section presents the experimental setup where the BBA_CO as a text feature selection optimizer is tested on two standard corpora of English text datasets, two Malay datasets, and one Arabic corpus. The pre-processing process, classifier and evaluation metrics used are also presented.

4.1. Pre-Processing

The pre-processing tasks are employed before the BBA_CO is tested on the text feature selection problem that involve normalization, tokenization, stop words removal and removal of the less frequent words. The normalization process removes non-letters and punctuation marks. Then, the capital letters are converted to small letters. In the tokenization process, the documents are divided into terms. The stop words (words without discriminative meaning) are then removed. Next, the words that rarely appear in the whole corpus are removed as they seem not to be significant to the classification process.

4.2. The Dataset

Three standard datasets including two English corpora namely Reuters-21578 and WebKB, two Malay corpora namely Mix-DS and Harian-Metro and one Arabic corpus namely, Al-Jazeera news are used in order to assess the performance of the proposed approach.

4.2.1. Reuters-21578 Dataset

This dataset contains 21,578 text files, which were collected from the Reuters newswire. These files were non-uniformly divided into 135 classes. This work utilizes the top 10 classes namely earn, acquisition, trade, ship, grain, crude, interest, money-fx, corn and wheat, which contain 4808 documents.

4.2.2. WebKB Dataset

This dataset is a collection of web pages from four different college websites, contains 8282 web pages assigned to 7 classes. In this work, only four classes are used as in the literature, which contains 2803 documents. The utilized classes are student, faculty, course and project.

4.2.3. Mix-DS

This dataset was manually collected from several websites. The total number of documents in this dataset is 12,269, distributed unevenly among 6 categories.

Table 2. Number of collected documents from each website for the classes of Mix-DS.

	Sukan	Bisnes	Pendidikan	Sains-Teknologi	Hiburan	Politik	Total
utusan online	170	238	300	220	1030	0	1958
mstar	3392	10	0	0	4533	0	7935
astroawani	96	96	0	96	96	96	480
bharian	527	1292	25	0	27	25	1896
Total	4185	1636	325	316	5686	121	12,269

shows the number of documents in each class and the websites where the documents were collected.

4.2.4. Harian Metro Dataset

The Harian Metro dataset has been collected from the Harian Metro website, and it consists of 7920 documents, distributed evenly (720 documents) among 11 categories namely; Sukan, Bisnes, Pendidikan, Teknologi, Hiburan, Dekotaman, Global, Vroom, Sihat, Sanati and Addin.

4.2.5. Al-Jazeera News Dataset

This dataset consists of 1500 text documents distributed equally among five categories (Economy, Science, Politics, Sport, and Art) and each category has 300 text documents. This dataset was collected from the Al-Jazeera news channel website (www.aljazeera.net) (accessed on 26 March 2022).

In this work, three classification algorithms are used in the experiments, which are Support Vector Machine (SVM), Naïve Bayes (NB) and K-Nearest Neighbor (KNN).

4.3. Evaluation Metric

In this work, the proposed methods are evaluated internally and externally. The internal evaluation concerns itself with the evaluation of the feature selection method, such as the quality and diversity of the population. On the other hand, the external evaluation evaluates the resulting feature set when utilizing it for classification. Although multiple evaluation metrics are utilized for evaluation, the computational cost is not considered as it is one time cost. Thus, the evaluation focuses on the quality of the resulted feature set and the classification results.

4.3.1. Internal Evaluation Metric

For the internal evaluation of the text-feature selection method presented in this work, different evaluation metrics will be used. These metrics are:

Solution Quality

This is also called the fitness value or the objective value of the solution. This value is calculated using the adapted dependency degree measure of RST and the number of features. The following equation is used to evaluate each candidate solution:

Fitness(x_i) = p × dep(x_i) + (1 − p) × (1/size(x_i))

where x_i is the feature subset found by solution i. Fitness is calculated based on both the dependency measure of rough set theory (dep(x_i)), and the length of the feature subset (size(x_i)). p is a parameter that controls the relative weight of dependency value and feature subset length, where p ∈ [0, 1]. This formula denotes that the dependency value and feature subset size have a different effect on the evaluation. In this study, the dependency value is considered to be more important than the subset length, so p is set to 0.8, as in [69,70].

Size of the Selected Feature Set

This metric evaluated the reduction ability of the text-feature selection method. A good method should be able to produce a high-quality feature set with a smaller number of features. The number of features in the resulting feature set is compared with the original number of features before the feature selection process to evaluate the reduction rate.

Reduction Rate

This metric is also used to evaluate the reduction ability of the TFS method in percentage. The reduction rate is calculated based on the number of original features and the number of selected features as:

$$Reduction rate (\%)=100\times \frac{\#original features-\#selected features }{\#original features}$$

Diversity during the Search Process

The population’s diversity of different generations in the algorithm (i.e., generations 1, 20, 50, 80 and 100) is measured graphically to evaluate the exploration ability of the algorithm during the search process and its ability to maintain the population’s diversity to later generations.

Convergence Behavior

The convergence of the population is shown graphically to evaluate the ability of the method to keep improving the population and avoid premature convergence. The convergence of the whole population is shown in average, and the convergence of five randomly selected solutions is also shown to show their improvement progress during the search process.

Statistical Tests

The significance test (the t-test: two-sample assuming unequal variances) is conducted as a statistical analysis to compare the algorithms. The t-test is a statistical check of two population means. The t-test was successfully used for comparing two groups of results over multiple datasets for its simplicity, safety and robust results [71]. To perform the t-test, the t State and t Critical two tail values are calculated by Microsoft Excel software. The first group of results is considered significantly higher than the second group if the t State value is greater than the t Critical two tail. The second group of results is considered significantly higher than the first group if the t State value is less than -t Critical two tail. The difference between the two groups’ results is considered not significant if the t State value is in the interval [−t Critical two tail, t Critical two tail].

In this work, the t-test is used for internal evaluation to measure the diversity and quality of the population. The population diversity is measured using the standard deviation values of 32 populations generated. The t-test is also used to measure the average quality of the population. The population’s quality is also measured using Best Relative Error (BRE), Average Relative Error (ARE) and Worst Relative Error (WRE) of the populations, where lesser values represent a better quality of the population.

4.3.2. External Evaluation Metric

The external evaluation concerns the classification performance by employing the selected feature set. The classification performance is measured by multiple evaluation criteria, which are discussed in the following subsections.

Classification Performance

To evaluate the classification performance using the selected feature sets, two widely used performance measures are used namely Micro Average Fl and Macro Average F1. The Macro Average F1 measure depends on precision (P), recall (R) and F-measure, which are calculated for each class as follows:

P = a/(a + b)

P = a/(a + b)

$$F_1=\frac{2p\times r}{p+r}$$

where a is the number of documents correctly classified; b is the number of documents incorrectly classified and c is the number of documents in the class. Macro Average F1 is calculated as below:

$$F_1^{macro}=\frac{1}{m}{\displaystyle \sum }_{i=1}^m{F}_1\left(c_i\right)$$

where m is the number of classes and F1(c_i) is the F1 measure for the i^th class.

Micro Average F1 is calculated globally based on the global precision and recall. Calculations for precision and recall for micro averaging as given by [72] are shown below:

$$\begin{gathered} P^{\mu }=\frac{{\sum }_{i=1}^m{a}_i}{{\sum }_{i=1}^m\left(a_i+b_i\right)} \\ {R}^{\mu }=\frac{{\sum }_{i=1}^m{a}_i}{{\sum }_{i=1}^m\left(a_i+c_i\right)} \end{gathered}$$

where a, b, c, m are the same variables used in the previous equations of precision (P) and recall (R), and μ indicates Micro Averaging. Micro Average F1 is calculated as follows:

$$F_1^{Micro}=\frac{2\times p^{\mu }\times R^{\mu }}{p^{\mu }\times R^{\mu }}$$

Statistical Test

The t-test is conducted as a statistical analysis to compare two groups of classification results. If the difference in the results is above a certain value, this indicates that the text-feature selection method is significantly efficient.

5. Results and Discussion

In this work, the performance of the proposed approach is measured based on the internal and external evaluation metrics as discussed in Section 4.3.

5.1. Internal Evaluation

The performance of the BBA_CO is compared to BBA_LHS in terms of population diversity, convergence behavior and the solution quality. Note that BBA_LHS is a binary Bat algorithm that is modelled based on one population, contrary to the BBA_CO where it is modelled as a multi-population binary Bat algorithm. In addition, please note that the modified LHS is used to generate initial population(s) for both BBA_LHS and BBA_CO.

5.1.1. Population Diversity

Population diversity is represented in the form of a distribution of solutions during the optimization process (i.e., at the 1st, 20th, 50th, 80th and 100th generations). Figure 6 and Figure 7 show the distribution of the solutions for the Reuters and WebKB datasets, respectively. It can be seen that the diversity of the population is well controlled by the BBA_CO in comparison to BBA_LHS where poor diversity can be noted after the first quarter of the search process, contrary with the BBA_CO where the solutions in the population are fairly distributed across the feature space.

5.1.2. Convergence Behavior

A comparison between the results achieved by BBA_CO and BBA_LHS is depicted in graphical form to show the convergence behavior as shown in Figure 8 and Figure 9 on the Reuters and WebKB datasets, respectively.

It can be noted that the population of BBA_CO converged slower than the population of BBA_LHS due to controlling diversity during the search process. However, the average quality of the solutions at the end of the optimization in BBA_CO is higher than in BBA_LHS. From the results of the convergence, it is clear that in BBA_LHS, the population converge faster and stagnate in the first third of the search process, while in BBA_CO the convergence is slower and the stagnation occurs by the last third of the search process. As a conclusion, BBA_LHS initialization could be the best choice if the purpose was improving the whole population within a small number of generations, while BBA_CO is better if the purpose is controlling the diversity and obtaining a better final solution. Figure 10 and Figure 11 compare the convergence of five randomly selected solutions using BBA_LHS and BBA_CO in Reuters and WebKB datasets, respectively.

5.1.3. Statistical Test

The t-test is conducted for multiple groups of results to compare BBA_LHS and BBA_CO in terms of the population’s diversity and quality in different generations. The population’s diversity is measured using standard deviation, which is greater for more diverse populations. Moreover, the population’s quality is measured by the average quality of its solutions and the relative errors as shown in the following subsections.

Standard Deviation

The quality dispersion of the solutions in the population is measured on the basis of the standard deviation (SD), which is higher for more diverse distributions and lower for less diverse distributions. SD is calculated as in the equation below. Consequently, the SD of the population in certain generations (1, 20, 50, 80, 100) is calculated and recorded for 32 different runs in BBA_LHS and BBA_CO for each dataset.

Table 3. The t-test of standard deviation of population in different generations using BBA_CO and BBA_LHS.

Dataset	Reuters		WebKB
Method	BBACO	BBALHS	BBACO	BBALHS
	generation 1
Mean	0.16	0.14	0.14	0.13
p-Value	0.13	0.36	0.74	0.29
t State	3.75		4.38
t Critical two-tail	2.00		2.01
	generation 20
Mean	0.15	0.12	0.14	0.13
p-Value	0.58	0.16	0.36	0.12
t State	13.96		1.94
t Critical two-tail	2.03		2.01
	generation 50
Mean	0.14	0.08	0.11	0.09
p-Value	0.12	0.18	0.53	0.13
t State	7.93		3.61
t Critical two-tail	2.01		2.00
	generation 80
Mean	0.12	0.06	0.09	0.08
p-Value	0.20	0.29	0.18	0.08
t State	15.67		2.98
t Critical two-tail	2.01		2.00
	generation 100
Mean	0.10	0.05	0.07	0.06
p-Value	0.17	0.13	0.09	0.07
t State	11.72		2.24
t Critical two-tail	2.01		2.02

shows the results of the t-test which is conducted between the results obtained using BBA_LHS and BBA_CO to investigate whether or not they are statistically different:

$$SD=\sqrt{\frac{{\sum }_{i=1}^n{\left(fitnes{s}_i-\overline{fitness}\right)}^2}{n}}$$

where SD stands for standard deviation; n is the population size; fitness_i is the fitness of solution i; $\overline{fitness}$ is the average fitness of the initial population.

Table 3 shows that the SD of the population in the selected generations for both datasets using BBA_CO is significantly higher than the SD when BBA_LHS is used. Additionally, it could be stated that when using the t-test, the first group of results is significantly higher than the second group of results when t State is greater than t Critical two-tail. On the other hand, if t State is less than −t Critical two-tail, this means that the results in the second group are significantly higher than the results in the first group. Meanwhile, if t State is within the interval [−t Critical two-tail, t Critical two-tail], the difference is not considered to be significant.

The results in Table 3 show that t State is higher than t Critical two-tail in all the selected generations. The value of t State is higher than t Critical two-tail in all generations. The results of the t-test indicate that the population diversity which is expressed by standard deviation, is higher in BBA_CO during the search process. This reveals the ability of BBA_CO to control the population’s diversity during the search process.

Population Quality

In order to measure the population quality, the average quality of the solutions will be used. This measure was utilized by the authors of [73] to measure the population quality. It is defined as the average quality of solutions in the population, as given in the following equation:

$$population quality (\%)=\left(1-\frac{ \overline{fitness}-bes{t}_{known}}{bes{t}_{known}}\right)$$

where $\overline{fitness}$ is the average quality (i.e., dependency value; fitness) of solutions in the population; $bes{t}_{known}$ is the best-known quality that the solution could reach.

The population quality in certain generations (1, 20, 50, 80, 100) is calculated and recorded for 32 different runs in BBA_LHS and BBA_CO for each dataset.

Table 4. The t-test of population quality in different generations using BBA_CO and BBA_LHS.

Dataset	Reuters		WebKB
Method	BBACO	BBALHS	BBACO	BBALHS
	generation 1
Mean	37.28	44.66	32.08	45.86
p-Value	0.39	0.35	0.26	0.31
t State	−7.60		−18.59
t Critical two-tail	2.02		2.00
	generation 20
Mean	52.33	55.36	56.48	64.28
p-Value	0.32	0.16	0.13	0.36
t State	−3.26		−7.38
t Critical two-tail	2.04		2.04
	generation 50
Mean	60.23	57.54	64.15	70.78
p-Value	0.64	0.06	0.19	0.15
t State	2.14		−8.38
t Critical two-tail	2.00		2.04
	generation 80
Mean	72.98	70.83	73.09	71.47
p-Value	0.43	0.13	0.16	0.29
t State	3.48		2.58
t Critical two-tail	2.03		2.03
	generation 100
Mean	73.98	71.16	74.27	72.79
p-Value	0.51	0.29	0.22	0.18
t State	2.79		2.82
t Critical two-tail	2.00		2.01

shows the results of the t-test which is conducted for average population quality that are obtained using BBA_LHS and BBA_CO on the Reuters and WebKB datasets.

Table 4 shows that the average population quality in the early stages of the search process (i.e., generation 1 and 20) is significantly better with BBA_LHS in both datasets. By the second half of the search process (i.e., generation 50) the average population quality with BBA_CO improved and statistically outperformed the quality with BBA_LHS, in the Reuters dataset. In WebKB, the average population quality obtained by BBA_CO statistically outperformed the quality obtained by BBA_LHS by the last quarter of the search process (i.e., generation 80). In both datasets, the quality of the population in the end of search process is statistically better with BBA_CO than BBA_LHS.

From the results of the statistical test, it is clear that the population with BBA_CO improves more slowly than that of BBA_LHS. However, the population with BBA_CO converges with better quality than that of BBA_LHS. The reason behind the slow improvement of the population with BBA_CO is the controlling of diversity by the cooperative coevolving strategy. In this way, the algorithm is allowed to continue exploring a wider range of the search space, which significantly improves the final solution.

Relative Errors

The relative errors include Best Relative Error (BRE), Worst Relative Error (WRE) and Average Relative Error (ARE). BRE, WRE and ARE can be used to measure how far the distance between the best and worst solutions and the average quality of the population from the best-known quality that could be achieved by the solution. Whenever the error rate of a solution decreases, the solution becomes closer to the optimum solution. BRE, WRE and ARE have been used to compare BBA_CO with BBA_LHS in certain generations (i.e., 1, 20, 50, 80, 100). The t-tests of BRE in the selected generations with BBA_LHS and BBA_CO are shown in

Table 5. The t-test of BRE in different generations using BBA_CO and BBA_LHS.

Dataset	Reuters		WebKB
Method	BBACO	BBALHS	BBACO	BBALHS
	generation 1
Mean	35.83	31.96	47.16	34.30
p-Value	0.63	0.39	0.08	0.15
t State	2.11		18.56
t Critical two-tail	2.03		2.03
	generation 20
Mean	24.03	20.50	23.75	23.50
p-Value	0.79	0.38	0.13	0.15
t State	3.14		0.21
t Critical two-tail	2.03		2.01
	generation 50
Mean	20.42	20.00	21.86	23.37
p-Value	0.72	0.25	0.42	0.27
t State	0.45		−1.04
t Critical two-tail	2.03		2.00
	generation 80
Mean	14.80	18.50	18.98	21.50
p-Value	0.54	0.29	0.36	0.16
t State	−7.24		−3.22
t Critical two-tail	2.01		2.04
	generation 100
Mean	14.60	18.50	15.89	21.00
p-Value	0.84	0.27	0.45	0.19
t State	−4.57		−7.05
t Critical two-tail	2.02		2.03

.

Observing the values of t State and t Critical two-tail in Table 5, it is clear that the BRE of the populations of BBA_CO are significantly higher in the beginning of the search process, in both datasets. By the end of the first half (i.e., generation 50), the difference between BRE in the two groups of results reveals no significance in both of the datasets. By the later generations of the search process (i.e., generations 80 and 100), the BRE with BBA_CO reduced and became significantly less than that of BBA_LHS in both datasets. The results of the statistical test for BRE indicate that in the beginning of the search process, the best solution in BBA_LHS outperforms that of BBA_CO. Then, with the progress of the search process, the best solution of BBA_CO improves until it outperforms that of BBA_LHS. This shows that BBA_CO is less subjected to stagnation than BBA_LHS. The t-test of ARE in the selected generations with BBA_LHS and BBA_CO are shown in

Table 6. The t-test of ARE in different generations using BBA_CO and BBA_LHS.

Dataset	Reuters		WebKB
Method	BBACO	BBALHS	BBACO	BBALHS
	generation 1
Mean	90.85	86.90	90.72	86.78
p-Value	0.14	0.38	0.40	0.16
t State	7.45		3.50
t Critical two-tail	2.00		2.02
	generation 20
Mean	87.91	86.25	88.00	85.24
p-Value	0.18	0.21	0.32	0.06
t State	1.88		2.51
t Critical two-tail	2.03		2.00
	generation 50
Mean	84.73	85.61	84.00	81.65
p-Value	0.09	0.19	0.10	0.13
t State	−0.90		1.86
t Critical two-tail	2.00		2.00
	generation 80
Mean	78.07	68.95	75.12	79.15
p-Value	0.17	0.32	0.07	0.16
t State	6.88		−2.00
t Critical two-tail	2.02		2.00
	generation 100
Mean	74.52	62.95	70.85	77.64
p-Value	0.24	0.14	0.21	0.08
t State	5.64		−2.84
t Critical two-tail	2.00		2.02

.

As shown in Table 6, the ARE of BBA_CO population is significantly higher than that of BBA_LHS at the beginning of the search process in both datasets. In later generations (i.e., generations 20 and 50), the ARE of the population of BBA_CO keeps reducing but the difference with the other group (i.e., ARE of the population of BBA_LHS) is not statistically significant, in the Reuters dataset. By the last quarter of the search process (i.e., generations 80 and 100), the ARE of the population of BBA_CO became statistically lower than that of BBA_LHS, in the Reuters dataset. In the WebKB dataset, the ARE of BBA_CO was significantly higher than that of BBA_LHS within the first half of the search process. Later, the ARE of the BBA_CO population keeps reducing, but the difference with the other group is not significant. The results of ARE are consistent with those of BRE, as the population in BBA_CO improves slower than that of BBA_LHS because of controlling diversity. The T-test of WRE in the selected generations with BBA_LHS and BBA_CO is shown in

Table 7. The t-test of WRE in different generations using BBA_CO and BBA_LHS.

Dataset	Reuters		WebKB
Method	BBACO	BBALHS	BBACO	BBALHS
	generation 1
Mean	90.85	86.90	90.72	86.78
p-Value	0.14	0.38	0.40	0.16
t State	7.45		3.50
t Critical two-tail	2.00		2.02
	generation 20
Mean	87.91	86.25	88.00	85.24
p-Value	0.18	0.21	0.32	0.06
t State	1.88		2.51
t Critical two-tail	2.03		2.00
	generation 50
Mean	84.73	85.61	84.00	81.65
p-Value	0.09	0.19	0.10	0.13
t State	−0.90		1.86
t Critical two-tail	2.00		2.00
	generation 80
Mean	78.07	68.95	75.12	79.15
p-Value	0.17	0.32	0.07	0.16
t State	6.88		−2.00
t Critical two-tail	2.02		2.00
	generation 100
Mean	74.52	62.95	70.85	77.64
p-Value	0.24	0.14	0.21	0.08
t State	5.64		−2.84
t Critical two-tail	2.00		2.02

.

The results in Table 7 indicate that WRE is significantly higher in the populations of BBA_CO until the end of the search process in the Reuters dataset. In the WebKB dataset, WRE remains significantly higher with BBA_CO than BBA_LHS within the first quarter of the search process. Then, the difference of WRE between both populations of BBA_CO and BBA_LHS became not significant until later generations. By the end of the search process, WRE of the BBA_CO population became significantly higher than that of BBA_LHS population. The reason behind that is that the populations of BBA_CO are more diverse, and thus, the solutions are more distributed in the search space until the later generations. In contrast, the solutions of BBA_LHS population in the later generations are found to be closer to each other. In this way, the quality of the worst solution with BBA_LHS remains better than the one with BBA_CO in Reuters dataset.

The Quality and Reduction Rate of the Selected Feature Set

This subsection compares BBA_CO and BBA_LHS in terms of the quality of the final solution (i.e., the resultant feature set), the size of the selected feature set and the reduction rate.

Table 8. Quality of the final solution, number of features and reduction rate of BBA_CO in the Reuters dataset.

Class Name	Quality of Final Solution	# of the Selected Features	Reduction Rate (%)
Earn	0.87	307	82.87
Acquisition	0.88	394	84.76
Trade	0.90	103	89.24
Ship	0.88	56	84.09
Grain	0.73	26	82.31
Crude	0.75	131	85.11
Interest	0.83	45	88.64
Money-fx	0.76	78	87.62
Corn	0.74	89	88.35
Wheat	0.79	157	81.06

and

Table 9. Quality of the final solution, number of features and reduction rate of BBA_CO in WebKB dataset.

Class Name	Quality of Final Solution	# of the Selected Features	Reduction Rate (%)
Student	0.81	226	90.23
Faculty	0.87	179	93.39
Course	0.74	105	94.41
Project	0.71	103	93.51

compare the two methods based on the mentioned metrics in the Reuters and WebKB datasets, respectively.

Table 8 and Table 9 clearly show that BBA_CO improves the quality and the reduction rate of the selected feature set compared with the BBA_LHS results. The improvement is likely attributed to the ability of BBA_CO to combine the advantages of dividing the solutions into smaller parts and to maintain the populations’ diversity. Dividing the solution into smaller solutions allows BBA_CO to better optimize the solution components (i.e., parts) resulting in a better final solution. The cooperative coevolving strategy in BBA_CO directs the algorithm to better convergence.

5.2. External Evaluation

This section compares the classification results obtained by the feature sets generated by BBA_CO and BBA_LHS for the Reuters and WebKB datasets. In order to compare the classification performance using BBA-based methods, the classification is also conducted using Chi-Square (CHI), Information Gain (IG) and Gini Index (GI) feature-selection methods, which were successfully used in the literature for TFS [74,75,76]. Micro Average F1 and Macro Average F1 are used as performance measures for text classification. Some of the best classification results are shown in

Table 10. Classification results using the feature sets generated by BBA_LHS and BBA_CO.

Dataset	Metric	Classifier	CHI	IG	GI	BBALHS	BBACO
Reuters	Micro average F1	NB	88.80	90.80	90.50	92.78	93.76
		SVM	86.70	89.40	89.80	92.55	94.08
		KNN	86.30	89.90	90.30	92.63	93.17
	Macro average F1	NB	79.50	78.50	77.20	89.87	90.03
		SVM	81.70	77.20	75.90	88.76	90.05
		KNN	66.60	68.30	69.10	88.04	89.49
WebKB	Micro average F1	NB	79.50	78.20	77.50	91.79	92.72
		SVM	88.30	89.30	89.10	91.64	92.84
		KNN	65.30	66.70	65.70	90.94	92.06
	Macro average F1	NB	78.20	76.90	75.90	89.82	91.67
		SVM	87.00	87.90	87.80	89.37	90.51
		KNN	60.70	62.50	61.40	88.03	90.87

comparing BBA_CO with BBA_LHS, CHI, IG and GI.

Table 10 shows the classification results of the Reuters and WebKB datasets using three classifiers (i.e., NB, SVM and KNN) in Micro Average F1 and Macro Average F1. The results clearly show the improvement in the classification performance with all classifiers when using BBA-based TFS methods. In addition, BBA_CO improves the classification performance over BBA_LHS as a result of the improved feature sets selected by BBA_CO.

Table 11. Precision (P), Recall (R) and F-measure (F) of each class in the Reuters dataset using the BBA_LHS.

Class	NB			SVM			KNN
	P	R	F	P	R	F	P	R	F
Earn	99.18	98.60	98.89	98.09	95.35	96.70	94.02	92.25	93.13
Acquisition	93.16	98.60	95.80	85.95	96.78	91.04	90.54	82.68	86.43
Trade	96.36	91.98	94.12	97.05	87.97	92.29	87.17	92.04	89.54
Ship	97.56	78.43	86.96	96.76	75.29	84.69	82.25	88.39	85.21
Grain	91.96	71.53	80.47	94.20	78.14	85.42	81.02	89.03	84.84
Crude	81.70	94.58	87.67	87.76	82.41	85.00	97.79	91.19	94.37
Interest	95.08	94.79	94.93	96.46	88.02	92.05	82.32	87.12	84.65
Money-fx	97.37	79.20	87.35	91.10	81.99	86.31	89.20	98.67	93.70
Corn	89.63	82.59	85.97	90.15	86.23	88.15	87.87	78.35	82.84
Wheat	88.76	84.50	86.58	88.63	83.39	85.93	89.67	82.12	85.73
Average	93.08	87.48	89.87	92.61	85.56	88.76	88.18	88.18	88.04

,

Table 12. Precision (P), Recall (R) and F-measure (F) of each class in the WebKB dataset using BBA_LHS.

Class	NB			SVM			KNN
	P	R	F	P	R	F	P	R	F
Project	96.97	85.69	90.98	83.91	97.75	90.30	88.68	97.38	92.83
Course	92.18	86.14	89.06	91.56	83.66	87.43	90.21	84.93	87.49
Faculty	84.85	91.16	87.89	88.98	82.51	85.62	85.96	76.44	80.92
Student	89.48	93.33	91.36	98.42	90.19	94.13	98.11	84.62	90.87
Average	90.87	89.08	89.82	90.72	88.53	89.37	90.74	85.84	88.03

,

Table 13. Precision (P), Recall (R) and F-measure (F) of each class in the Reuters dataset using BBA_CO.

Class	NB			SVM			KNN
	P	R	F	P	R	F	P	R	F
Earn	94.31	92.66	93.48	96.54	91.41	93.90	91.74	93.94	92.83
Acquisition	90.87	97.25	93.95	89.17	98.36	93.54	91.77	96.56	94.10
Trade	94.22	91.18	92.68	96.06	91.28	93.61	94.59	88.91	91.66
Ship	97.56	78.43	86.96	93.62	77.39	84.73	95.37	70.55	81.10
Grain	94.92	77.78	85.50	93.73	82.64	87.84	89.13	76.47	82.32
Crude	96.49	92.12	94.25	95.16	91.57	93.33	93.99	95.54	94.76
Interest	93.29	89.87	91.55	91.53	88.87	90.18	93.29	93.86	93.57
Money-fx	93.80	90.81	92.28	92.07	87.69	89.83	95.62	92.49	94.03
Corn	90.31	79.52	84.57	90.64	82.59	86.43	86.89	82.98	84.89
Wheat	82.87	87.45	85.10	88.26	85.98	87.11	86.97	84.35	85.64
Average	92.86	87.71	90.03	92.68	87.78	90.05	91.94	87.56	89.49

and

Table 14. Precision (P), Recall (R) and F-measure (F) of each class in the WebKB dataset using BBA_CO.

Class	NB			SVM			KNN
	P	R	F	P	R	F	P	R	F
Project	95.88	92.29	94.05	96.67	91.57	94.05	91.76	93.35	92.55
Course	87.72	93.57	90.55	97.18	89.33	93.09	93.55	86.74	90.02
Faculty	88.61	79.89	84.02	81.13	88.22	84.53	85.57	92.17	88.75
Student	97.69	98.42	98.05	86.85	94.20	90.38	90.79	93.58	92.16
Average	92.48	91.04	91.67	90.46	90.83	90.51	90.42	91.46	90.87

show the classification results in terms of Precision, Recall and F-measure for the Reuters and WebKB datasets using BBA_LHS and BBA_CO, respectively.

Table 13 and Table 14 clearly show the improvement in the classification results when using BBA_CO as a TFS method. Comparing the classification results achieved by BBA_CO with those of BBA_LHS (i.e., Table 11 and Table 12), it is found that the average F is improved with the three classifiers in both datasets. From the experimental results, it is shown that dividing the solutions into smaller ones and optimizing each part then improved the resulting final solution, controlled the population’s diversity and improved the classification performance as a result of the improved selected feature set. In order to test if the performance of BBA_CO is significantly better than that of BBA_LHS, the t-test needs to be applied to their classification results. The results of the statistical test are presented in

Table 15. The t-test of classification results of BBA_CO vs. BBA_LHS methods.

Dataset	Reuters		WebKB
Method	BBALHS	BBACO	BBALHS	BBACO
	NB		NB
Mean	92.25	93.41	91.35	92.22
p-Value	0.08	0.18	0.49	0.33
t State	−8.23		−5.83
t Critical two-tail	2.02		2.00
	KNN		KNN
Mean	92.17	93.09	90.59	91.88
p-Value	0.10	0.59	0.62	0.99
t State	−6.37		−9.74
t Critical two-tail	2.00		2.02
	SVM		SVM
Mean	91.83	93.67	91.20	92.14
p-Value	0.56	0.59	0.42	0.98
t State	−16.54		−7.41
t Critical two-tail	2.00		2.03

.

As shown in Table 15, the value of t State is less than -t critical two-tail in all cases. These results indicate that the second group of the classification results is significantly higher than the first group (i.e., the classification results obtained by using the selected feature set by BBA_LHS). The reason behind that is attributed to the cooperative coevolving strategy that is utilized in BBA_CO, which improves the performance of the text-feature selection method by dividing the solution into smaller ones with a smaller dimension than the original full solution. This way, optimizing each part separately improves the performance and controls the diversity of the population.

5.3. Results on Non-English Datasets

This section investigates the ability of the proposed text-feature selection method to be generalized for different languages. Therefore, the two variants of the proposed text-feature selection method i.e., BBA_LHS and BBA_CO, are tested on two Malay and one Arabic datasets. The quality of the selected feature sets, the number of the selected features and reduction rate are reported in this section. Furthermore, the classification results utilizing three classifiers (i.e., NB, SVM and KNN) are reported and discussed.

5.3.1. Results of Malay Datasets

The classification results of Malay datasets utilizing three classifiers and the two versions of the proposed method are reported in this subsection. The results are also compared with the classification results using no feature selection method and using Chi-Square feature selection method. Chi-Square is a well-known method that has been successfully utilized for feature selection [9,74,77]. Firstly, the characteristics of the resulting feature sets by BBA_LHS and BBA_CO are displayed in

Table 16. Characteristics of the selected feature sets in Mix-DS, where Q indicates the quality of feature set, #F indicates to number of features, R indicates reduction rate (%).

Class	Q-BBALHS	Q-BBACO	#F	#F-BBALHS	#F-BBACO	R-BBALHS	R-BBACO
Bisnes	0.55	0.63	4785	1058	1014	77.89	78.81
Hiburan	0.56	0.60	10,357	1468	1328	85.83	87.18
Pendidikan	0.63	0.71	2212	564	428	74.50	80.65
Politik	0.54	0.54	1534	341	325	77.77	78.81
Sains-Teknologi	0.59	0.65	3023	784	751	74.07	75.16
Sukan	0.61	0.67	8320	1127	946	86.45	88.63

and

Table 17. Characteristics of the selected feature sets in Harian Metro dataset, where Q indicates the quality of feature set, #F indicates to number of features, R indicates reduction rate (%).

Class	Q-BBALHS	Q-BBACO	#F	#F-BBALHS	#F-BBACO	R-BBALHS	R-BBACO
Addin	0.416	0.483	3925	315	256	91.97	93.48
Bisnes	0.536	0.547	2564	335	184	86.93	92.82
Dekotaman	0.564	0.58	3396	358	314	89.46	90.75
Global	0.667	0.638	2052	259	237	87.38	88.45
Hiburan	0.645	0.651	2947	356	330	87.92	88.80
Pendidikan	0.655	0.685	3284	365	303	88.89	90.77
Santai	0.517	0.592	3940	428	402	89.14	89.80
Sihat	0.662	0.688	3943	382	374	90.31	90.51
Sukan	0.406	0.489	2041	363	327	82.21	83.98
Teknologi	0.537	0.669	2818	201	194	92.87	93.12
Vroom	0.616	0.668	3150	362	342	88.51	89.14

for Mix-DS and Harian Metro dataset, respectively. Then, the classification results are reported in

Table 18. Classification results achieved by using the original dataset, the feature sets selected by Chi-square, BBA_LHS and BBA_CO.

Dataset	Metric	Classifier	No FS	Chi-Square	BBALHS	BBACO
Mix-DS	Micro average F1	NB	76.32	84.32	88.50	89.34
		SVM	76.01	84.6	88.04	88.72
		KNN	75.94	84.07	87.77	88.24
	Macro average F1	NB	73.40	83.58	86.61	88.42
		SVM	74.72	82.15	86.98	87.38
		KNN	73.35	82.42	86.35	87.47
Harian-Metro dataset	Micro average F1	NB	68.20	78.94	82.38	83.56
		SVM	67.38	79.19	82.16	83.71
		KNN	66.77	78.51	81.64	83.18
	Macro average F1	NB	67.49	76.09	81.10	83.61
		SVM	66.29	77.13	80.86	82.16
		KNN	66.02	75.84	80.55	81.98

.

Table 16 and Table 17 compare the performance of BBA_LHS and BBA_CO in terms of quality, number of selected features and reduction rate. The tables show that both methods are able to reduce the dimensionality of the feature space efficiently. In addition, it is clear that BBA_CO outperforms BBA_LHS in terms of the quality of the feature set and the reduction rate. Table 18 reports the classification results using the selected feature sets by BBA_CO compared with the selected sets by Chi-square, BBA_LHS and with the original feature set. The impact of using an efficient feature selection method is clear, which is consistent with what has been concluded in the literature. It is clear how Chi-square has improved the classification results over the original dataset. However, the classification performance with BBA_LHS and BBA_CO is clearly improved over Chi-square. The results demonstrate the ability of BBA-based methods to select discriminative feature sets. Furthermore, using a coevolutionary technique has clearly improved the performance of the TFS method, which has been reflected in the improved classification results.

Table 19. Classification results of Mix-DS using NB, SVM and KNN classifiers in terms of precision (P), recall (R) and F-measure (F).

Class	NB-No FS			SVM-No FS			KNN-No FS			NB-Chi-Square			SVM-Chi-Square			KNN-Chi-Square
	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F
Bisnes	0.93	0.79	0.85	0.91	0.80	0.85	0.93	0.78	0.85	0.93	0.91	0.92	0.91	0.89	0.90	0.95	0.93	0.94
Hiburan	0.81	0.99	0.89	0.82	0.93	0.87	0.59	0.83	0.69	0.88	0.96	0.92	0.97	0.96	0.96	0.83	0.86	0.84
Pendidikan	0.93	0.43	0.59	0.87	0.49	0.62	0.86	0.56	0.68	0.59	0.82	0.69	0.64	0.75	0.69	0.76	0.87	0.81
Politik	0.94	0.54	0.68	0.94	0.62	0.75	0.83	0.64	0.72	0.89	0.83	0.86	0.81	0.73	0.77	0.83	0.96	0.89
Sains-Teknologi	0.65	0.41	0.50	0.59	0.43	0.50	0.67	0.69	0.68	0.64	0.65	0.64	0.62	0.64	0.63	0.73	0.48	0.58
Sukan	0.99	0.82	0.90	0.96	0.84	0.89	0.85	0.73	0.78	0.98	0.98	0.98	0.97	0.98	0.98	0.82	0.93	0.87
Average	0.87	0.66	0.73	0.85	0.68	0.75	0.79	0.71	0.73	0.82	0.86	0.84	0.82	0.83	0.82	0.82	0.84	0.82
Class	NB-BBALHS			SVM-BBALHS			KNN-BBALHS			NB-BBACO			SVM-BBACO			KNN-BBACO
	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F
Bisnes	0.99	0.89	0.94	0.99	0.89	0.93	0.88	0.93	0.90	0.98	0.93	0.95	0.96	0.94	0.95	0.89	0.89	0.89
Hiburan	0.77	0.99	0.86	0.73	0.94	0.82	0.74	0.89	0.81	0.92	0.89	0.90	0.93	0.89	0.91	0.89	0.86	0.88
Pendidikan	0.94	0.73	0.82	0.96	0.75	0.84	0.79	0.92	0.85	0.78	0.91	0.84	0.77	0.90	0.83	0.80	0.88	0.83
Politik	0.99	0.82	0.90	0.99	0.80	0.89	0.85	0.95	0.89	0.86	0.96	0.91	0.83	0.96	0.89	0.90	0.81	0.85
Sains-Teknologi	0.81	0.77	0.79	0.76	0.93	0.84	0.83	0.72	0.77	0.83	0.70	0.76	0.82	0.66	0.73	0.78	0.88	0.83
Sukan	0.94	0.85	0.89	0.94	0.86	0.90	0.95	0.97	0.96	0.94	0.95	0.95	0.93	0.95	0.94	0.97	0.97	0.97
Average	0.91	0.84	0.87	0.89	0.86	0.87	0.84	0.89	0.86	0.88	0.89	0.88	0.87	0.88	0.87	0.87	0.88	0.87

and

Table 20. Classification results of Harian Metro dataset using NB, SVM and KNN classifiers in terms of precision (P), recall (R) and F-measure (F).

Class	NB-No FS			SVM-No FS			KNN-No FS			NB-Chi-Square			SVM-Chi-Square			KNN- Chi-Square
	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F
Addin	0.93	0.57	0.70	0.80	0.57	0.66	0.45	0.69	0.55	0.98	0.73	0.84	0.71	0.96	0.82	0.94	0.76	0.84
Bisnes	0.62	0.76	0.69	0.80	0.55	0.65	0.86	0.50	0.63	0.68	0.78	0.73	0.88	0.65	0.75	0.84	0.64	0.73
Dekotaman	0.88	0.51	0.64	0.77	0.53	0.62	0.93	0.53	0.68	0.98	0.68	0.80	0.65	0.88	0.75	0.66	0.73	0.70
Global	0.93	0.50	0.65	0.65	0.64	0.64	0.95	0.52	0.67	0.81	0.74	0.77	0.73	0.81	0.77	0.87	0.66	0.75
Hiburan	0.91	0.50	0.64	0.76	0.52	0.62	0.93	0.51	0.66	0.90	0.69	0.78	0.90	0.58	0.70	0.69	0.85	0.76
Pendidikan	0.86	0.60	0.71	0.74	0.66	0.69	0.88	0.63	0.73	0.65	0.73	0.69	0.76	0.65	0.70	0.78	0.69	0.73
Santai	0.81	0.51	0.63	0.56	0.55	0.55	0.86	0.54	0.66	0.75	0.86	0.80	0.66	0.71	0.68	0.62	0.82	0.71
Sihat	0.87	0.57	0.69	0.54	0.79	0.64	0.42	0.79	0.54	0.74	0.62	0.68	0.75	0.87	0.81	0.87	0.76	0.81
Sukan	0.59	0.98	0.74	0.66	0.91	0.77	0.59	0.97	0.73	0.56	0.72	0.63	0.97	0.76	0.85	0.74	0.99	0.85
Teknologi	0.78	0.64	0.70	0.85	0.58	0.69	0.78	0.66	0.71	0.98	0.68	0.81	0.91	0.73	0.81	0.88	0.69	0.77
Vroom	0.63	0.63	0.63	0.89	0.63	0.74	0.91	0.55	0.69	0.98	0.76	0.85	0.93	0.77	0.84	0.86	0.59	0.70
Average	0.80	0.62	0.67	0.73	0.63	0.66	0.78	0.63	0.66	0.82	0.73	0.76	0.80	0.76	0.77	0.80	0.74	0.76
Class	NB-BBALHS			SVM-BBALHS			KNN-BBALHS			NB-BBACO			SVM-BBACO			KNN-BBACO
	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F	P	R	F
Addin	0.84	0.77	0.80	0.77	0.86	0.81	0.82	0.76	0.79	0.71	0.95	0.81	0.90	0.82	0.86	0.86	0.84	0.85
Bisnes	0.77	0.94	0.85	0.87	0.77	0.82	0.85	0.76	0.80	0.89	0.73	0.80	0.82	0.78	0.80	0.84	0.79	0.82
Dekotaman	0.92	0.76	0.83	0.91	0.81	0.86	0.94	0.81	0.87	0.89	0.92	0.90	0.75	0.86	0.80	0.83	0.87	0.85
Global	0.78	0.88	0.82	0.72	0.86	0.78	0.75	0.85	0.80	0.85	0.71	0.77	0.84	0.77	0.80	0.79	0.84	0.81
Hiburan	0.73	0.79	0.76	0.83	0.73	0.78	0.71	0.80	0.76	0.87	0.92	0.89	0.77	0.89	0.83	0.86	0.82	0.84
Pendidikan	0.84	0.89	0.86	0.84	0.78	0.81	0.88	0.89	0.88	0.81	0.85	0.83	0.75	0.87	0.81	0.88	0.77	0.82
Santai	0.71	0.77	0.74	0.74	0.81	0.78	0.71	0.85	0.77	0.93	0.77	0.84	0.89	0.76	0.82	0.82	0.79	0.80
Sihat	0.87	0.79	0.83	0.82	0.79	0.81	0.76	0.76	0.76	0.76	0.87	0.81	0.88	0.92	0.90	0.81	0.81	0.81
Sukan	0.79	0.83	0.81	0.89	0.74	0.81	0.74	0.88	0.81	0.89	0.92	0.90	0.90	0.74	0.81	0.88	0.74	0.80
Teknologi	0.88	0.73	0.80	0.84	0.79	0.82	0.83	0.79	0.81	0.84	0.78	0.81	0.83	0.80	0.81	0.88	0.81	0.84
Vroom	0.84	0.79	0.81	0.91	0.76	0.83	0.88	0.76	0.81	0.78	0.85	0.81	0.72	0.89	0.80	0.84	0.73	0.78
Average	0.82	0.81	0.81	0.83	0.79	0.81	0.81	0.81	0.81	0.84	0.84	0.84	0.82	0.83	0.82	0.84	0.80	0.82

show the classification results in terms of precision, recall and F-measure for Mix-DS and Harian Metro dataset, respectively.

5.3.2. Results of Arabic Dataset

The classification results of the Arabic dataset utilizing three classifiers are reported in this subsection. The quality of the selected feature sets, the number of the selected features and reduction rate are also reported in

Table 21. Characteristics of the selected feature sets in Al-Jazeera news dataset, where Q indicates the quality of feature set, #F indicates number of features, R indicates reduction rate (%).

Class	Q-BBALHS	Q-BBACO	#F	#F-BBALHS	#F-BBACO	R-BBALHS	R-BBACO
Economy	0.54	0.61	1541	537	397	65.15	74.24
Politics	0.59	0.64	1427	343	152	75.96	89.35
Sport	0.62	0.68	1874	294	199	84.31	89.38
Science	0.63	0.68	1505	235	176	84.39	88.31
Art	0.66	0.71	1693	544	267	67.87	84.23

.

Table 21 compares the performance of BBA_LHS and BBA_CO in terms of quality, number of selected features and reduction rate. It is noticeable that both methods were efficient in reducing the feature set. It is also clear that BBA_CO outperforms BBA_LHS with all classes, which indicates the efficiency of the proposed coevolutionary technique.

The classification results in terms of average precision, average recall and Macro average F1 are reported and compared with the state-of-the-art results in Table 22. The state-of-the-art studies in the TFS field utilized different datasets, classifiers and evaluation measures. However, the experimental setting such as the utilized database, classifiers and evaluation metrics should be same in order to reach fair comparison. Thus, few studies are comparable with the proposed method for using the same experimental setting. The methods used for comparison are:

• Binary Particle Swarm Optimization with k-nearest neighbor (BPSO-KNN) [78];
• Enhanced Genetic Algorithm (EGA) [69];
• Category relevant feature measure (CRFM) [79].

The authors of [78] combined Binary PSO and k-nearest neighbor to select a feature set for Arabic text classification. They used three Arabic datasets and three classifiers to evaluate the performance of their method. A feature selection method based on Enhanced Genetic Algorithm (EGA) was proposed in the study by [69]. Their method was evaluated using three Arabic datasets and two classification algorithms namely NB and AC. The same datasets and classification algorithms were used in the study of [79] which proposed three enhanced feature selection methods. However, the results of CRFM have been used in the comparison as it performs better than the other two methods. In the current study, the classification was performed using NB and SVM that where successfully utilized in previous studies for text classification [69,78,79,80].

Table 22. Classification results of BBA_R, BBA_LHS, and BBA_CO on Al-Jazeera news dataset compared with state-of-the-art results.

Metric	Classifier	BPSO-KNN	EGA	CRFM	BBALHS	BBACO
Macro Average precision	NB	85.76	91	88.77	91.34	91.64
	SVM	93.7	-	-	90.86	92.22
Macro Average Recall	NB	84.34	90.66	88.33	90.42	91.09
	SVM	92.98	-	-	91.58	92.88
Macro Average F1	NB	84.63	90.83	88.55	90.73	91.24
	SVM	93.12	-	-	91.07	92.42

Table 22. Classification results of BBA_R, BBA_LHS, and BBA_CO on Al-Jazeera news dataset compared with state-of-the-art results.

Metric	Classifier	BPSO-KNN	EGA	CRFM	BBALHS	BBACO
Macro Average precision	NB	85.76	91	88.77	91.34	91.64
	SVM	93.7	-	-	90.86	92.22
Macro Average Recall	NB	84.34	90.66	88.33	90.42	91.09
	SVM	92.98	-	-	91.58	92.88
Macro Average F1	NB	84.63	90.83	88.55	90.73	91.24
	SVM	93.12	-	-	91.07	92.42

Table 22 reports the classification results of the two variants of BBA-based TFS methods compared with three state-of-the-art results. Based on Table 22, it is obvious that the results of NB with BBAco outperform the results of all other methods. On the other hand, the results of SVM with BPSO-KNN outperform the results of the other methods. However, unlike BPSO-KNN, the proposed BBA-based methods likely perform above 90% with both utilized classifiers. Hence, it is clear that the proposed BBA-based TFS methods are efficient with the Arabic dataset.

Table 23. Classification results of Al-Jazeera news dataset using NB and SVM in terms of precision (P), recall (R) and F-measure (F).

Class	NB-BBALHS			SVM-BBALHS			NB-BBACO			SVM-BBACO
	P	R	F	P	R	F	P	R	F	P	R	F
Economy	94.85	83.65	88.90	91.13	95.20	93.12	89.04	91.51	90.26	94.17	99.59	96.80
Politics	92.28	97.32	94.73	85.25	97.55	90.99	86.92	97.85	92.06	88.50	97.63	92.84
Sport	91.10	85.55	88.24	93.87	85.98	89.75	97.55	88.90	93.02	92.66	83.98	88.11
Science	90.94	89.11	90.02	86.30	87.50	86.90	91.05	86.72	88.83	86.44	90.50	88.42
Art	87.51	96.46	91.77	97.73	91.69	94.61	93.62	90.46	92.01	99.34	92.69	95.90
Average	91.34	90.42	90.73	90.86	91.58	91.07	91.64	91.09	91.24	92.22	92.88	92.42

shows the classification results of the Al-Jazeera news dataset with BBA_LHS and BBAco in terms of precision, recall and F-measure.

Table 23 shows that the average precision, recall and F-measure with BBA_CO are clearly higher than those of BBA_LHS in terms of average precision, recall and F-measure in all cases. These results demonstrate the efficiency of the proposed TFS method with the Arabic dataset. This successfully shows that the proposed BBA-based text-feature selection method is applicable with the Arabic language.

In order to test whether the improvement of BBA_CO over BBA_LHS is significant or not, the statistical test (i.e., the t-test) is performed to the classification results of both methods. The statistical test compares the results of BBA_LHS with those of BBA_CO using NB and SVM classifiers. The t-test is conducted for the results of BBA_LHS vs. BBA_CO, to test whether the difference is significant or not. The results of the t-test are reported in

Table 24. The t-test of classification results of BBA_LHS vs. BBA_CO methods.

	BBALHS	BBACO
	NB
Mean	90.78	91.23
p-Value	0.20	0.96
t Stat	−6.89
t Critical two-tail	2.00
	SVM
Mean	91.11	92.52
p-Value	0.07	0.46
t Stat	−23.04
t Critical two-tail	2.00

.

As shown in Table 24, the value of t State is less than -t Critical two-tail with both classifiers. These results indicate that the second group of classification results (i.e., the classification results obtained using the selected feature set by BBA_CO) is significantly higher than the first group (i.e., the classification results obtained using the selected feature set by BBA_LHS). These results demonstrate the efficiency of BBA_CO to improve the resulting feature set, and therefore improve the classification performance by utilizing the coevolutionary technique as explained above in order to generating a better feature set.

Observing the results of the five utilized datasets in this study, namely Reuters, WebKB, Mix-DS, Harian Metro dataset and Aljazeera news dataset, it is noticed that the proposed text-feature selection method has a similar effect to all datasets, regardless of the language. The reason behind that is mostly attributed to the nature of BBA-based TFS methods, which depend on the mathematical calculations and do not consider the semantic attributes. Therefore, the performance of the proposed methods is independent from the dataset language.

6. Conclusions

Feature selection is a crucial step in text classification in order to overcome the high dimensionality of the feature space and improve the classification accuracy. Thus, many feature selection methods have been proposed in the literature. Those methods could be ranking methods, which rank the features and select the top ranked ones, or meta-heuristic-based methods that work as a wrapper around certain classification algorithms. However, the ranking methods ignore the correlation between features, while the wrappers are classifier-dependent. To overcome those limitations, this paper introduced the cooperative binary Bat algorithm (BBA_CO) and investigated its performance for the text feature selection. The quality of the final solution and the number of selected features are also compared with those of BBA_LHS; in order to test the impact of cooperative BBA.

The text classification was performed on two slandered English datasets, namely Reuters and WebKB, to evaluate the discriminative ability of the produced feature set using BBA_CO. The best classification results obtained with BBA_CO were 94% and 92.8% in terms of Micro Average for the Reuters and WebKB datasets, respectively. In comparison, the best results obtained by the other methods were 90.5% and 89.3% for the Reuters and WebKB datasets, respectively, and those results were obtained by IG. The text classification was also performed for two Malay and one Arabic datasets. The statistical test demonstrated that the improvement of the classification performance was significant. The experimental results in this work have shown the ability of the proposed method to improve the final solution, control the population’s diversity and improve the text classification accuracy. In addition, the proposed method was approved to be general in terms of dataset language.

For future work, the proposed coevolutionary method could be tested on a different population-based meta-heuristic algorithms to evaluate its performance. Additionally, the proposed coevolutionary method could be adapted to different high dimensional optimization problems. Moreover, designing a feature selection method based on parallel BBA could also be a better option for very high-dimensional datasets.