A systematic literature review on meta-heuristic based feature selection techniques for text classification

Research questions

This SLR is purposed to define the guidance gained from the previous studies using the MH techniques for FS in text classification. Table 1 presents five research questions discussed in this SLR. From the previous studies, MH techniques that have been used for FS were identified in (RQ1). The analysis of these studies has been conducted to answer the questions. RQ2 determines which FS-MH techniques have been used for text classification. RQ3 compares the performance of MH-FS techniques with the traditional techniques. The purpose of this question is to discover if MH methods are superior to conventional methods. RQ4 identified the advantages and limitations of different MH techniques to guide the selection of appropriate MH techniques, while the final question (RQ5) investigated how RSS can be used as FS.

Search strategy and study selection

Detailed analysis of this step is described in four subsections which are search terms, literature sources, search process, and study selection.

Search terms

Five steps were conducted to extract the search terms which are as follows (Malhotra, 2015):

1. Extract the key terms from the research questions.

2. Determine the synonyms and alternative spellings for the main terms.

3. Explore the keywords and terminology from existing research articles.

4. Combine the synonyms and alternative spellings using the Boolean operator “OR”.

5. Connect the main terms using the Boolean operator “AND”.

All research terms were derived from the explored topic. These terms are feature selection, attribute selection, text classification, text categorization, meta-heuristics, and metaheuristics. The final search terms that connect with the Boolean operators were as follows: ((“feature selection” OR “attribute selection”) AND (“text classification” OR “text categorization”) AND (“meta-heuristics” OR “metaheuristics”)).

Literature sources

The relevant studies were investigated through the Scopus, Science Direct, and Google Scholar databases. The records identified in the period from 2008 to 2022 used “feature selection” and “attribute selection” as the main keywords and the rest of the keywords to specify and limit the selected studies.

Search process

The number of studies obtained from the Scopus, Science Direct, and Google Scholar databases using the main keywords were as follows: 45,236 for Scopus, 53,030 for Science Direct, and 19,300 for Google Scholar. These studies were then further filtered using specific keywords such as “text classification” or “text categorization,” resulting in 5,186 studies for Scopus, 1,922 studies for Science Direct, and 842 studies for Google Scholar. Another set of specific keywords, namely “meta-heuristic” or “metaheuristic,” yielded 226 studies for Scopus, 438 studies for Science Direct, and 34 studies for Google Scholar. After the initial search, the relevant studies were chosen according to predefined inclusion and exclusion criteria explained in the following subsection.

Study selection

The studies obtained were further narrowed down to include only articles published between 2015 and 2022. The resulting numbers were 194 studies for Scopus, 200 studies for Science Direct, and 25 studies for Google Scholar. The Inclusion-Exclusion criterion was then applied to limit the search scope. The purpose was to evaluate all selected studies that either facilitated or directly addressed at least one research question in the field. Additionally, the analysis focused on articles and review articles, and only English articles were included. Based on these criteria, the list of studies was reduced to 143 for Scopus, 38 for Science Direct, and 18 for Google Scholar. This initial list was further analyzed and filtered by examining the titles, keywords, and abstracts to remove irrelevant papers. The systematic literature review (SLR) was conducted on research published between 2015 and 2022. The final list consisted of 91 studies for Scopus, 15 for Science Direct, and five for Google Scholar. Furthermore, any duplicate papers from Science Direct and Google Scholar were removed, resulting in a final list of 91, 12, and five studies for all databases. In addition, there are 15 selected studies were chosen out of limitation. Table 2 presents the details of the inclusion and exclusion criteria.

Quality assessment criteria

The selected studies were assessed with great care and seriousness to maintain a high-quality standard. This assessment included evaluating the novelty of the proposed techniques and the ability of the studies to address at least one research question. Furthermore, special consideration was given to choosing high-quality studies from high-impact journals available in the Scopus, Science Direct, and Google Scholar digital libraries. These measures were taken to ensure a thorough quality check and maintain the overall standard of the selected studies.

Data extraction

During the data extraction phase, the selected studies were utilized to gather crucial information necessary to tackle the research questions. The extracted information encompassed various aspects such as author details, publication year, applied meta-heuristic (MH) techniques, utilized datasets, employed classifiers, performance measures used for evaluation, and the obtained results. This collected information was then organized and analyzed to facilitate further data synthesis.

Data synthesis

The data synthesis phase in the SLR involves summarizing and interpreting the collected information from the selected studies. This phase aims to address the research questions through analysis, discussion, and various forms of representation such as tables, graphs, charts, etc. The SLR processes are typically executed multiple times to ensure an effective review process that yields the most relevant and suitable studies. The process of this particular SLR began with the identification of research questions, followed by the application of relevant search terms on the Scopus, Science Direct, and Google Scholar digital libraries. This initial search yielded 143, 38, and 18 research papers, respectively. After analyzing and filtering the studies based on inclusion and exclusion criteria, as well as removing any duplicated papers, the final number of obtained studies was 91, 12, and five, a total of 108 articles. Figure 2 illustrates the search steps protocol using the PRISMA flowchart, highlighting the progression from the initial search to the final selection of studies. After thoroughly searching through the selected studies, a total of 123 studies were found to be valuable for this SLR, as they exhibited a high level of relevance in addressing the chosen research domain.

Overview of the selected study

After conducting a scan and filtering process on studies published between 2015 and 2022, a total of 108 relevant studies were initially obtained. However, 15 studies were out of limitations and were included. Therefore, the final number of studies included in this review amounted to 123. Figure 4 visually represents the distribution of these studies over the period from 2015 to 2022, indicating that the highest number of studies was observed in 2019, while the lowest number of studies occurred in 2017.

The quality of the selected studies was assessed by considering the quality and impact factors of the journals in which they were published. This ensured that high-quality studies were included. Table 3 presents a comprehensive list of the journals that published the selected studies, along with the corresponding number of studies in each journal. It also provides information on the Quartile ranking of these journals in the International Scientific Indexing (ISI) and their impact factors. Additionally, the Quartile ranking in Scopus is indexed with their Scientific Journal Ranking (SJR). It can be concluded that the ISI journals accounted for 85.19% of the total number of selected studies, further confirming the overall quality of the included literature.

The selected studies were assigned unique identifiers (IDs) for easy reference and consistency throughout the subsequent subsections. Table 4 presents a comprehensive list of all the studies included in the review, along with their corresponding IDs and references. Furthermore, Table 5 presents a summary of the selected studies’ IDs along with the research questions they have addressed. The analysis reveals that all the selected studies have contributed to answering RQ1. For RQ2, 28 studies have provided relevant insights. Similarly, 26 studies have addressed RQ3, while 40 studies have tackled RQ4. It is worth mentioning that one paper, which was outside the predefined limitations, addressed RQ5 by utilizing RSS to enhance SVM in text classification. More detailed information on these findings can be found in the subsequent subsections.

RQ1: which metaheuristic (MH) techniques have been utilized for feature selection (FS)?

In this section, the focus is on discussing and identifying the MH techniques that have been utilized for feature selection in various machine learning problems, including pattern recognition, email classification, microarray data classification, sentiment analysis, and text classification. The findings from the primary selected studies reveal that all the studies have provided insights into (RQ1). These studies have classified the MH techniques into three main groups according to their sources of inspiration. These groups are Evolutionary Algorithms (EA), Physics-Based (PB) Algorithms, and Swarm Intelligence (SI) Algorithms. This categorization provides a broad understanding of the different types of MH techniques employed for FS across various domains (Kumar & Bawa, 2020).

Evolutionary Algorithms (EAs) draw inspiration from the natural processes of evolution. One of the commonly used algorithms in this category is the Genetic Algorithm (GA) belongs to a class of optimization algorithms that draw inspiration from the principles of natural selection and genetics. They mimic the principles of evolution to solve complex problems by iteratively searching and refining a population of potential solutions, that has been utilized for feature selection in numerous studies. Specifically, in RP13, RP22, RP27, RP32, RP33, RP37, RP40, RP46, RP62, RP85, and RP106. Another algorithm in the EA category is Differential Evolution (DE), which is used to solve continuous optimization problems, and was utilized for feature selection in RP8 and RP55. Evolutionary Population Dynamics (EPD) combines concepts from evolutionary biology and population dynamics with computational methods, it was used in RP31, Imperialist Competitive Algorithm (ICA) draws inspiration from the socio-political behavior of imperialistic systems, it was applied in RP17 and RP88. RP95 utilized the Binary Coyote Optimization Algorithm (BCOA), which draws inspiration from the intelligent behavior exhibited by coyotes in their natural environment. The Golden Eagle Optimizer (GEO) inspired by the behavior and characteristics of golden eagles in nature, is used by RP102. Biogeography-based Optimization (BBO) is influenced by the principles of biogeography, which involves the study of the distribution of biological organisms across different geographic regions, it is used by RP107. These studies demonstrate the application of different evolutionary algorithms for feature selection in various machine-learning problems.

Secondly, is the Physics-Based (PBs) Algorithms, which aim to mimic physical rules in their search process. Several common algorithms in this category have been employed for feature selection in the selected studies. Simulated Annealing (SA) draws inspiration from the annealing process used in metallurgy, it was utilized in RP4, RP28, RP43, and RP87. The Harmony Search (HS) algorithm takes inspiration from the musical improvisation process, it was applied in RP2, RP6, RP13, RP71, and RP99. The Gravitational Search Algorithm (GSA) derives its inspiration from the fundamental law of gravity and the motion of celestial bodies, it was used in RP21, RP28, and RP68. The Teaching Learning Based Optimization (TLBO) algorithm draws inspiration from the teaching and learning processes that take place in a classroom, it was employed in RP29 and RP103. Water Cycle algorithms (WCA) were utilized in RP43. Atom Search Optimization (ASO) was applied in RP73. The multi-Verse Optimizer (MVO) algorithm inspired by the concept of multiple universes and parallel universes in theoretical physics, was used in RP69. The Interior Search Algorithm (ISA) designed for solving constrained optimization problems, was employed in RP74. Lastly, RP93 utilizes the Black Hole Optimization (BHO) algorithm, a robust stochastic optimization technique that takes inspiration from the behavior of black holes in outer space. The key distinction between EAs and PBs lies in the mechanism of communication and movement of search agents within the search space. PB algorithms rely on physical rules to guide the search process, while EAs are inspired by evolutionary processes. This difference in approach enables PB algorithms to explore the search space using physics-inspired mechanisms.

The third group consists of Swarm intelligence (SI) algorithms, which draw inspiration from the collective behavior observed in swarms, herds, flocks, or schools of living organisms in nature. While these algorithms share similarities with evolutionary algorithms (EAs) and population-based (PB) algorithms in terms of their mechanism, SI algorithms leverage the simulated collective and social intelligence of these creatures to guide the interactions among search agents. The number of newly proposed SI algorithms exceeds those of EAs and PBs. One widely adopted SI algorithm is Ant Colony Optimization (ACO), which is utilized in several selected studies as a popular feature selection (FS) technique (RP1, RP3, RP5, RP7, RP12, RP15, RP16, RP20, RP26, RP34, RP42, RP52, RP59, RP61, RP63, and RP86). Additionally, Particle Swarm Optimization (PSO) is another notable SI algorithm employed in RP14, RP36, RP41, RP50, RP63, RP65, RP67, and RP100. It is worth emphasizing that the realm of Swarm Intelligence (SI) has witnessed a significant influx of novel algorithms in recent times, substantially broadening the array of choices available for optimization tasks. Among the more recent SI algorithms highlighted in the selected studies, notable examples include the Artificial Bee Colony (ABC) algorithm, applied in RP8, RP24, and RP88. Additionally, the Antlion Optimization (ALO) algorithm has been employed in RP9, RP18, and RP54, while the Crow Search Algorithm (CSA) has been identified in RP44, RP47, and RP77. Furthermore, the Whale Optimization Algorithm (WOA) has been utilized in RP30, RP55, RP76, and RP97. Other noteworthy SI algorithms encompass the Cuckoo Search (CS) algorithm, which has been employed in RP39, RP64, RP83, and RP105. Additionally, the Grasshopper Optimization Algorithm (GOA) has been found in RP31, RP56, RP60, RP66, and RP96, and the Grey Wolf Optimization (GWO) algorithm has been utilized in RP18, RP47, RP49, RP58, RP70, RP96, and RP101. and the Mayfly (MF) algorithm which is used in RP71 is introduced as a novel technique for addressing FS problems. This innovative method takes a hybrid approach, synergizing the strengths found in traditional optimization techniques like PSO, GA, and FA. These recent SI algorithms have substantially augmented the repertoire of options available for optimization tasks, providing researchers and practitioners with an expanded toolkit to effectively address intricate problems.

Furthermore, a diverse array of captivating swarm intelligence (SI) algorithms has emerged, significantly expanding the range of optimization techniques. These encompass the firefly algorithms (FFA) as scrutinized in RP23 and RP78, the Moth-Flame Optimizer (MFO) employed in RP18 and RP58, the Slap Swarm Optimization (SSO) explored in RP35 and RP81, and the Brainstorm Optimization (BSO) utilized in RP59 and RP72. Additionally, the Bat Optimization Algorithm (BA) has been studied in RP38, RP90, and RP108, while the Water Wave Optimization (WWO) has been investigated in RP80. Noteworthy algorithms also include the Butterfly Optimization Algorithm (BOA) employed in RP48 and RP98, the Selfish Herd Optimizer (SHO) studied in RP75, the Social Spider Optimization (SSO) explored in RP57, and the Artificial Fish Swarm Algorithm (AFSA) examined in RP45. Moreover, the Shuffled Frog Leaping Algorithm (SFLA) has been discussed in RP51 and RP104, and the Symbiotic Organism Search (SOS) Algorithm has been utilized in RP53. Additionally, the Pigeon-inspired Optimization (PIO) algorithm has been studied in RP82, the Krill Herd Optimization (KHO) explored in RP84, and the Dandelion Algorithm (DA) utilized in RP91.

In addition, the Horse Herd Optimization Algorithm (HOA) was thoroughly investigated in RP92. RP94 employed the Dragonfly Algorithm (DA), which took inspiration from the natural behavior of dragonflies. These SI algorithms exemplify the ingenuity and diversity of drawing inspiration from various natural phenomena and collective behaviors. Each algorithm derives inspiration from distinct aspects of nature or collective behavior, with the shared goal of providing effective optimization solutions for diverse problem domains. By simulating the behavior of fireflies, moths, slaps, bats, water waves, butterflies, selfish herds, social spiders, artificial fish swarms, shuffled frogs, symbiotic organisms, pigeons, krill herds, dandelions, and horse herds, these algorithms strive to offer efficient optimization solutions for a wide spectrum of problem domains. The continuous advancement and exploration of such algorithms contribute to the ever-evolving field of optimization, offering promising avenues for addressing intricate optimization challenges across various domains.

Several metaheuristic (MH) techniques find inspiration from various mathematical theories and concepts. For instance, the Chaotic Optimization Algorithm (COA), applied in RP9, RP18, and RP44, draws influence from Chaos Theory. RP4 and RP11 utilize the Greedy Randomized Adaptive Search Procedure (GRASP), a metaheuristic algorithm tailored for solving combinatorial problems. GRASP involves the construction and local search phases in each iteration. The Binary Coordinate Ascent (BCA) algorithm, employed in RP10, takes inspiration from the well-known coordinate descent algorithm. Additionally, the Variable Neighborhood Search (VNS) technique (RP19, RP25) tackles global optimization and combinatorial optimization problems by modifying the neighborhood of the current solution during the search process in a systematic manner. RP79 employs the Iterated Greedy (IG) technique, which addresses challenging combinatorial optimization problems through two phases: destruction and construction. The Sine Cosine Algorithm (SCA), utilized in RP85 and RP87, emulates the behavior of sine and cosine functions to uncover optimal solutions for optimization problems. RP89 utilizes the Jaya Optimization Algorithm (JOA), which draws inspiration from the Sanskrit concept of “Jaya,” signifying success or victory. This algorithm iteratively improves a population of solutions to discover an optimal or nearly optimal solution.

Table 6 presents a concise overview of the metaheuristic (MH) techniques utilized in the selected studies to address Research Question 1, focusing on Feature Selection (FS). These techniques have been utilized in the selected studies to tackle FS challenges and provide solutions in the context of Research Question 1.

Statistical analysis

The statistical analysis plan has been developed to assess the significance of differences between MH models in Table 6, which involves calculating summary statistics such as odds ratios (OR) for meta-analysis. OR for the data provided in Table 6 were calculated by constructing a 2 × 2 contingency table for each MH technique compared to the reference group (ACO). The ACO MH technique was designated as the reference group because it was the most frequently used technique. Then, the odds of studies using each technique were compared to the odds of studies using the reference technique (ACO). The odds ratio was calculated by dividing the number of studies using each technique by the number of studies using ACO as shown in Eq. (1). Then, the results were organized into Table 7.

(1) $$Oddsratio\left(eachtechniquevs.ACO\right)={\displaystyle \frac{Oddsofstudieswitheachtechnique}{OddsofstudieswithACO}}$$

RQ2: in the context of text classification, which specific MH techniques have been applied for FS?

In this section, the aim is to identify the metaheuristic-feature selection (MH-FS) techniques utilized in the selected studies for text classification. Additionally, we provide an overview of the commonly used datasets, classifiers, and performance evaluation metrics in the context of MH-FS for text classification machine learning techniques. As depicted in Table 5, a total of 28 studies (RP2, RP15, RP17, RP19, RP20, RP36, RP37, RP41, RP42, RP43, RP45, RP49, RP55, RP58, RP64, RP69, RP78, RP79, RP83, RP84, RP87, RP88, RP92, RP93, RP96, RP97, RP100, and RP108) have employed MH techniques for FS in text classification. These studies offer insights into the application of MH-FS techniques, highlighting the datasets commonly used, the classifiers employed, and the frequently utilized performance evaluation metrics in the domain of text classification machine learning.

RQ2.1: what are the datasets employed in the application of MH-FS for text classification?

Different datasets have been employed for MH-FS in text classification, depending on the specific classification task, such as sentiment analysis, spam classification, or general text classification. Additionally, the choice of datasets has been influenced by the language used, including English, Arabic, and Turkish. In Table 8, ten selected studies (RP15, RP20, RP42, RP55, RP58, RP69, RP79, RP83, and RP88) utilized sentiment analysis for text classification. All of these studies used the English language, except for RP55, which employed Arabic. Similarly, five studies (RP2, RP36, RP43, RP84, and RP92) focused on spam text classification. Among the studies that performed English text classification, RP17, RP19, RP41, RP45, RP64, RP49, RP78, RP37, RP93, RP96, RP97, RP100, and RP108 were identified, while RP49 and RP78 also incorporated Arabic text classification, and RP37 included Turkish text classification.

RQ2.2: which classifiers have been used with MH-FS in text classification?

The text classification process encompasses three primary stages: text preprocessing, feature selection (FS), and constructing a text classification model with a machine learning classifier to evaluate the performance of different FS techniques. The selected studies have utilized different classifiers for FS-MH in text classification. Table 9 presents eight classifiers along with their definitions and the studies where they were applied for text classification. These classifiers include support vector machine (SVM), naïve Bayesian (NB), k-nearest neighbor (KNN), decision tree (DT), multilayer perceptron (MLP), artificial neural networks (ANN), centroid based algorithm (CBA), and AdaBoost. As shown in Table 9, SVM, NB, and KNN are the most commonly used classifiers in text classification, followed by DT and MLP, with AdaBoost being used to a lesser extent. The least utilized classifiers are ANN and CBA. Figure 5 provides a visual representation of the number of studies employing different classifiers in text classification.

RQ2.3: which performance evaluation metrics are commonly utilized to assess the effectiveness of MH-FS in text classification?

Among the studies that have been reviewed in the field of MH-FS for text classification, a range of evaluation metrics have been utilized. These metrics serve to evaluate and compare the performance of diverse models developed through a variety of machine learning and statistical methods. The metrics in question include Precision, Recall, F-measure, Accuracy, AUC (area under the curve), the number of selected features, and Stability. Table 10, detailed the descriptions and definitions of these evaluation metrics, along with information on which studies made use of them. Figure 6 has been prepared to offer a visual overview of the usage frequency of these metrics. As shown in Fig. 6, Precision and Recall are the most commonly employed metrics, with F-measure and Accuracy following closely behind. AUC and the count of selected features are also frequently employed for evaluation purposes. In contrast, Stability is a less commonly utilized metric in the analyzed studies.

RQ3: is there empirical evidence indicating that MH-FS techniques outperform traditional FS methods in the domain of text classification?

According to the selected studies, the comparison was done based on two types of studies which are: (1) studies compared MH-FS and traditional FS techniques and (2) studies compared MH-FS techniques with existing MH-FS techniques. In this section, the focus is on comparing MH with the traditional one. The MH-FS techniques perform better than the traditional FS In this review, 26 articles extracted from the chosen studies (RP1, RP2, RP3, RP5, RP6, RP8, RP10, RP12, RP13, RP17, RP23, RP26, RP30, RP31, RP34, RP41, RP42, RP46, RP50, RP58, RP60, RP61, RP62, RP69, RP78, and RP79) undertake comparisons of their methodologies with various traditional FS techniques such as information gain, ReliefF, Laplacian score (L-score), Fisher score (F-score), relevance–redundancy feature selection (RRFS), random subspace method (RSM), Chi-square, mutual Information, symmetrical-uncertainty, minimal-redundancy–maximal-relevance (mRMR), sequential feature selection (SFS), correlation-based feature selection (CFS), and so many other techniques.

To provide a scientific explanation without the need to delve into detailed descriptions of all the techniques, two examples representing the mentioned technologies have been presented such as RP1 and RP2. These two studies were chosen to ensure a clear and comprehensive understanding for the reader without being overwhelmed with additional details. In RP1, the RRFS-ACO MH technique consistently outperformed traditional techniques, demonstrating superior results when compared to information gain, gain ratio, symmetrical-uncertainty, Gini index, F-score, term variance (TV), L-score, mRMR, mutual correlation (MC), RSM, and RRFS. In RP2, the global best harmony-oriented harmony search (GBHS) achieved superior results when compared to conventional methods, including Chi-square, feature selection based on comprehensive measures, t-test-based feature selection, information gain using term frequency, and an improved term frequency-inverse document frequency approach. Table 11 lists the most frequently used traditional FS techniques along with their definitions and their studies.

RQ4: what are the discernible strengths and weaknesses of MH techniques in the context of FS?

This section identifies and summarizes the strengths and weaknesses of metaheuristic (MH) techniques as reported by researchers. it is focusing on the strengths and weaknesses that have been supported by multiple studies. MH techniques have demonstrated strong performance in feature selection (FS) problems. They are particularly praised for their ability to effectively handle redundant features and high-dimensional data. Table 12 summarizes the strengths of MH techniques based on the selected studies, along with the studies that support each strength. Concurrently, Table 13 summarizes the weaknesses of MH techniques, accompanied by the studies that provide evidence for each weakness. In summary, it is important to note that different MH techniques have varying advantages, and there is no universal solution for MH-FS techniques that fits all scenarios.

According to the previous discussion it expected that some different solutions that can efficiently improve the MH-FS which are as follows: use MH with the binary system, integrate with traditional FS techniques, integrate with some mathematics theories such as Chaos theory and Rough Set theory, and combine two MH techniques to take benefit from their advantages to complement each other. In addition, there are several reasons which inspire FS techniques to adopt MH techniques as highlighted by the research articles in RP3, RP4, RP6, RP42, RP53, RP57, and RP66. The reasons that motivate researchers to adopt MH in FS are summarized in Fig. 7. Briefly, there are various aspects of MH for FS following section highlights their strengths and discusses each aspect.

Aspects of MH based on FS

The various strengths of MH approaches for feature selection, including their robust search strategies, global optimization capabilities, parallelizability, adaptability, innovation potential, and cross-domain applicability, make them a powerful choice for addressing feature selection challenges across a wide range of applications. Researchers and practitioners often turn to MH when seeking efficient and effective solutions for feature subset selection.

1. Search strategy: MH approaches are known for their powerful search strategies. They excel at navigating complex, high-dimensional search spaces to find the optimal subset of features. These strategies balance exploration and exploitation to efficiently explore the space while exploiting promising solutions. This dynamic search process is a significant strength of MH, enabling them to find feature subsets that traditional methods might overlook due to their deterministic nature.

2. Global optimization: One of the key strengths of MH approaches is their ability to perform global optimization. In FS, it is crucial to find the best combination of features that leads to optimal model performance. MH can efficiently search for solutions across the entire feature space, which traditional methods may struggle to do. This global exploration capability is particularly valuable when dealing with large and complex feature sets.

3. Parallelization and scalability: MH approaches often lend themselves well to parallelization. This means they can be distributed across multiple processors or machines to expedite the search process, making them scalable for large datasets and high-dimensional feature spaces. This scalability is crucial for handling big data and complex applications, further showcasing the practicality of MH feature selection.

4. Adaptability: Another advantage of MH is its adaptability to various optimization objectives and constraints. Researchers can easily customize the objective function to reflect the specific goals of FS problem, such as maximizing classification accuracy, minimizing model complexity, or considering trade-offs.

5. Exploration of new techniques: MH approaches encourage the development and exploration of new techniques. Researchers continually innovate and propose novel metaheuristic algorithms specifically designed for FS. For example, the RSS may represent such an innovation. This innovation-driven aspect of MH contributes to the field’s dynamism and evolution.

6. Cross-domain applicability: MH approaches are not limited to a single domain. They can be applied to FS in various fields, not just text classification. This cross-domain applicability demonstrates the versatility and effectiveness of these techniques in solving FS problems in diverse application areas.

RQ5: how can the RSS be effectively leveraged as an FS technique?

RSS, which stands for Ringed Seal Search, is a metaheuristic (MH) technique introduced in reference (Saadi et al., 2016). It draws inspiration from the natural behavior of seal pups when selecting a secure lair to evade predators. In comparison to other MH techniques like genetic algorithms (GA) and particle swarm optimization (PSO), RSS demonstrates faster convergence towards the global optimum and maintains a better trade-off between exploration and exploitation (Saadi et al., 2016). Although RSS has not been widely employed as a feature selection (FS) technique according to the existing literature, it possesses the capability to optimize SVM parameters. Consequently, this optimization leads to enhanced classification accuracy when compared to traditional SVM approaches (Sharif et al., 2019). The RSS algorithm is primarily inspired by the search behavior of seal pups seeking optimal lairs to evade predators. It adopts a similar approach where the algorithm continually searches for better solutions and moves towards them. In the context of RSS, these “lairs” correspond to problem-specific representations, and the algorithm aims to optimize the quality or fitness of these representations. By iteratively improving the representations, RSS strives to identify the best possible solution. The scenario begins once the female seal gives birth to a pup in a birthing lair that has been created for this purpose. The seal pup technique entails of searching for and choosing the ideal lair by conducting a randomized walk to discover a new lair. The seal pup’s random walk alternates between the normal and urgent search modes because seals are sensitive to outer noise produced by predators. The pup’s normal mode is an intensive search among closely spaced lairs, which is described by Brownian walk. In urgent state, the pup leaves the proximity area and implements extensive search to discover new lair from scatter targets; this movement is described as Levy walk. The change among these two modes is realized by the random noise released via predators. The algorithm holds changing between normal and urgent modes until the global optimum is reached.

RSS is especially based on seal pup search for optimal lairs to escape predators. Each time a new lair that has perfect quality is found; the pup moves into it. In the end, the lair (habitat) with the optimal fitness (quality) it is the term that RSS is going to optimize. The RSS concepts is represented on the following depictions.

• i)

  Each female seal gives birth to a single pup at a time, selecting a random habitat for the pup.

• ii)

  The seal pup randomly explores its ecosystem to locate a suitable lair for protection against predators.

• iii)

  The movement of the seal pup can be categorized into two states: Normal, where the search is focused and follows a Brownian walk, and Urgent, where the search is expansive and follows a Levy walk.

• iv)

  If the best-seen lair $L^{best,t}$ among the current set of lairs $K$ is superior in terms of fitness value compared to the best lair $L^{best,t-1}$ from the previous iteration, $L^{best}$ is updated to $L^{best,t-1}$. Otherwise, $L^{best}$ remains unchanged.

Over time, inferior lairs will be discarded, and the seals will continue to explore and move towards better lairs or chambers, leading to convergence towards good solutions. The RSS algorithm will be adapted for feature selection by using the following steps:

1. Input: Provide the initial number of lairs for the search.

2. Output: The algorithm aims to find the best lairs based on some evaluation criteria.

3. Initialization: Generate birthing lairs’ initial number: Initialize the lairs as L_1 = (f = 1, 2, 3,…, n), where n is the initial number of lairs.

4. While (Stopping criterion): Repeat the following steps until a specific stopping criterion is met. This criterion could be a maximum number of iterations, reaching a satisfactory solution, or other conditions specific to the problem being solved.

5. If noise = false: Check if the noise parameter is set to false. If so, perform a Brownian walk in the proximity to search for a new lair. A Brownian walk is a random process where the next step is determined by a random direction.

6. Else: If the noise parameter is set to true, expand the search for a new lair using a Levy walk. A Levy walk is a random process that allows for long-range exploration and can provide more global search capabilities.

7. Evaluation: Evaluate the fitness of every new lair and compare with previous: Assess the fitness of each newly generated lair using an appropriate evaluation function metrics such as accuracy, recall, precision, etc. Compare the fitness of these lairs with the previously evaluated lairs.

8. If $L^{\wedge }\left(best,T\right)>L^{\wedge }\left(best,T-1\right):$ Check if the fitness of the current best lair (( $L^{\wedge }(best,T$) is greater than the fitness of the previous best lair ( $L^{\wedge }(best,T-1)$).

9. Choose the new lair: If the fitness of the current best lair is greater, select it as the new best lair ( $L^{\wedge }best=L^{\wedge }\left(best,T\right)).$

10. Else: If the fitness of the current best lair is not greater, go to step 4 and continue the search.

11. Rank the lairs: Once the stopping criterion is met, rank the lairs based on their fitness evaluations.

12. End the loop.

13. Termination: Return the best feature subset found as the result of the feature selection.

RSS can be employed for feature selection by following the previous steps. Firstly, the problem should be defined, and the features must be represented appropriately. Secondly, a population of potential feature subsets is initialized, and their fitness is evaluated using a suitable metric. Thirdly, the RSS algorithm is applied iteratively, with the search space being explored, the fitness function being assessed, and the best solutions being exploited. A termination criterion is defined to determine when the algorithm should stop. Finally, the selected features can be extracted from the best solution after the algorithm finishes. Figure 8 summarizes the RSS feature selection algorithm.

Future research directions

In this SLR, future research direction could focus on the further refinement and development of existing MH techniques and investigate how this technique can be optimized to enhance their performance in feature selection. It was noted that there are still some MH techniques, such as RSS, which have not been extensively explored for FS despite their effectiveness. RSS employs two search states, normal and urgent, and dynamically switches between them until the optimal solution is reached. This balance between exploitation and exploration enables RSS to find global optima faster than other techniques. Moreover, RSS has shown high accuracy in text classification problems, making it a promising choice for FS. Therefore, as a future direction, the study suggests the utilization of RSS as an FS technique in this research.

Further considerations addressing the challenges in feature selection research based on MH techniques require a multifaceted approach. Researchers can develop novel algorithms suitable for handling large-scale datasets efficiently while maintaining robustness and scalability. Additionally, strategies for handling dynamic environments and evolving datasets must be devised to ensure the adaptability of feature selection methods over time. Improving the interpretability and explainability of MH-based feature selection models is essential for gaining insights into the decision-making process. Hybrid approaches combining MH with other optimization techniques, machine learning algorithms, binary systems, or mathematical theories such as chaos theory can leverage the strengths of each method, leading to enhanced performance and flexibility. Additionally, the limitations of MH techniques require targeted strategies tailored to each algorithm’s weaknesses. For instance, to overcome the challenges associated with ACO, efforts can focus on developing enhanced variants that improve efficiency and convergence speed, possibly through parameter tuning or hybridization with other optimization methods. Similarly, for HS, PSO, and ABC, optimizations could target convergence speed and space complexity by refining the search strategies or introducing adaptive mechanisms. DE’s instability and GA’s issues with crossover could be mitigated by incorporating diversity maintenance mechanisms or alternative operators. In addition, for algorithms like CSA and SSO, refining the balance between exploration and exploitation is crucial, possibly through algorithmic modifications or parameter adjustments. Finally, interdisciplinary research can uncover applications beyond text classification, while comparing MH techniques with deep learning methods can offer insights into their relative strengths and limitations.