The rapid evolution of Artificial Intelligence (AI), deep learning, and big data analytics has transformed modern science and industry, leading to an era defined by immense data complexity. In fields such as genomics, medical imaging, cybersecurity, and financial forecasting, datasets now contain thousands of feature attributes, signals, or measurements that collectively describe a phenomenon or system. While this abundance of information offers opportunities for powerful predictive modeling, it also introduces severe computational and analytical challenges. High-dimensional data often contain redundant, irrelevant, or noisy features that degrade model accuracy, increase training time, and complicate interpretability. As models grow deeper and data become larger, the need for efficient dimensionality reduction techniques becomes increasingly vital [1]. Feature selection has emerged as one of the most effective solutions to this problem. It seeks to identify a subset of the most informative features while discarding those that contribute little or no value to prediction or classification. Unlike feature extraction, which transforms data into new dimensions, feature selection preserves the original meaning of each variable, maintaining interpretability. By eliminating redundancy and focusing on relevance, feature selection improves learning efficiency, enhances generalization, and reduces overfitting, key requirements for dependable and explainable machine learning. This makes it indispensable in domains where decisions must be both accurate and transparent, such as medical diagnostics, credit scoring, industrial control, and environmental modeling [2].

However, identifying the optimal subset of features remains one of the most challenging problems in data analysis. The search space of possible feature combinations grows exponentially with the number of available features, making exhaustive enumeration computationally infeasible even for moderately sized datasets. This challenge belongs to the class of Nondeterministic Polynomial (NP)-hard problems, meaning that no known deterministic algorithm can guarantee an optimal solution within polynomial time. Traditional methods such as forward selection, backward elimination, or branch-and-bound approaches often fail to scale to modern data volumes. Consequently, researchers have turned to metaheuristic algorithms and stochastic optimization methods that efficiently approximate near-optimal solutions in vast, complex, and nonlinear spaces.

Metaheuristic algorithms have gained prominence in feature selection because they rely on intelligent sampling and adaptive search strategies rather than exhaustive enumeration. They operate through iterative exploration of candidate solutions, guided by randomized movements and feedback from the objective function. These algorithms are inherently flexible and require little prior knowledge of the problem structure. They balance two critical behaviors: exploration, which encourages diversity and prevents premature convergence to local optima, and exploitation, which intensifies search around promising regions. Achieving an effective equilibrium between these two processes is central to the success of any metaheuristic method [3], [4]. From 2020 to 2025, research in metaheuristic optimization has expanded dramatically. Hundreds of new algorithms have been introduced, inspired by mechanisms observed in nature, human society, and physical systems. This period represents a turning point in the evolution of metaheuristics, characterized by a shift from static, rule-based models toward dynamic, hybrid, and intelligent frameworks. The landscape now extends far beyond the classical algorithms of the past, such as genetic algorithms, particle swarm optimization, and ant colony optimization. Contemporary designs integrate concepts from behavioral psychology, swarm intelligence, physics, medicine, and even social interaction, creating a diverse taxonomy of optimization strategies [5].

These algorithms can be broadly grouped into four major categories. Evolutionary algorithms simulate the processes of biological evolution, using operators analogous to mutation, selection, and recombination to evolve solutions across generations. Physics-based algorithms draw on natural laws such as gravity, equilibrium, and thermodynamics to guide particle movement and energy transfer within a search space. Human–social algorithms imitate collective human behavior, learning, and cooperation, often incorporating decision-making, negotiation, or educational dynamics. Finally, swarm intelligence algorithms mimic the self-organized movement of social animals such as birds, fish, and insects, emphasizing distributed problem-solving and communication among agents [6].

Each of these categories contributes a unique perspective on optimization. Evolutionary algorithms excel at maintaining diversity through genetic operators, making them effective for multimodal and non-convex problems. Physics-based algorithms introduce deterministic mathematical formulations that ensure stable convergence and predictable dynamics. Human–social algorithms leverage cognitive and cooperative processes that adapt well to dynamic or imbalanced environments, such as data streams or evolving networks. Swarm intelligence algorithms, in contrast, are particularly suited for large-scale problems, as their decentralized nature promotes scalability and parallelism. Together, these families form a comprehensive toolkit for tackling the feature selection problem in both continuous and discrete domains [7].

Feature selection using metaheuristics has gained traction because these algorithms are naturally well-suited to the combinatorial nature of subset selection. Each feature subset can be represented as a binary vector, where each bit indicates whether a feature is included or excluded. The optimization objective is typically twofold: maximizing classification accuracy while minimizing the number of selected features. This dual objective can be addressed either by scalarizing both objectives into a single weighted function or by adopting multi-objective optimization frameworks that yield Pareto-optimal solutions. The latter approach has become increasingly popular because it offers greater flexibility and deeper insight into the trade-offs between accuracy and complexity [8].

The period from 2020 to 2025 has seen several innovations in the design and application of metaheuristics for feature selection. One major advancement has been the incorporation of chaotic systems and random perturbations to enhance population diversity and prevent stagnation. Chaos theory introduces deterministic randomness: small changes in initialization lead to large differences in trajectories, helping the algorithm escape local minima. Another significant direction has been the development of hybrid algorithms, combining two or more metaheuristics to exploit complementary strengths. For example, hybrid models may use swarm-based exploration to maintain diversity and evolutionary refinement to achieve convergence. Such combinations have yielded substantial improvements in accuracy and speed, especially for high-dimensional datasets [1].

Adaptive and self-regulating algorithms have also become a defining trend of the 2020–2025 period. Unlike static algorithms with fixed control parameters, adaptive metaheuristics adjust their parameters dynamically during the search process, guided by performance feedback or environmental cues. Reinforcement learning has been integrated into several designs to enable the algorithm to learn from its previous actions, improving operator selection and convergence strategy. This convergence between machine learning and optimization has created a new class of intelligent metaheuristics capable of self-improvement and cross-domain adaptability [8].

Swarm-based algorithms dominate recent research, accounting for more than half of newly proposed metaheuristics in the past five years. They are particularly attractive for feature selection because of their population diversity, distributed computation, and ease of binary adaptation. Popular examples include the Sparrow Search Algorithm, Marine Predators Algorithm, and Slime Mould Algorithm, each offering distinct exploration mechanisms modeled after animal behavior. Variants of these algorithms often incorporate adaptive inertia weights, chaotic maps, or opposition-based learning to improve performance on feature selection benchmarks. Their scalability has made them the preferred choice for applications in gene selection, medical image analysis, and cybersecurity, where high-dimensional datasets and computational efficiency are critical [9].

Human–social algorithms, on the other hand, have expanded the conceptual reach of metaheuristics beyond biological analogies. They replicate processes such as political negotiation, educational competition, and human memory retention. This class of algorithms has shown promise in handling data imbalance and dynamic decision spaces, as their structures naturally integrate elements of learning and feedback. They often introduce memory archives, role-based cooperation, or competition-driven selection, making them adaptive and interpretable, two qualities increasingly valued in machine learning applications [10].

Evolutionary and physics-based algorithms remain highly influential as well. Evolutionary models have evolved, incorporating psychological and sociological elements to model motivation, cooperation, and emotion, resulting in richer dynamics. Physics-based algorithms continue to offer mathematically elegant frameworks for simulating energy balance, gravitational fields, and molecular interactions to guide search trajectories. Binary and discrete adaptations of these algorithms have made them especially useful for feature selection, where decisions are fundamentally categorical.

Another defining characteristic of this era is the rise of hybrid and ensemble frameworks. Researchers increasingly integrate multiple algorithms or learning models into a single system to enhance robustness. Hybridization allows one algorithm to compensate for another’s weaknesses. For example, an evolutionary component may maintain global diversity, while a swarm component accelerates convergence. In some advanced implementations, metaheuristics are coupled with deep learning architectures, where optimization guides the selection of input features or hyperparameters for neural networks. This merging of optimization and learning blurs the boundary between searching and training, creating adaptive frameworks that evolve over time [11], [12].

Interpretability has become a parallel objective alongside accuracy and efficiency. In many real-world scenarios, particularly in medicine and finance, understanding why a feature subset is chosen is as important as achieving high predictive accuracy. Metaheuristic-based feature selection offers a natural path toward explainability, as it provides explicit binary decisions on which features to retain or remove. When combined with statistical or visualization tools, these results can reveal meaningful patterns, such as identifying biomarkers for disease or indicators for financial risk. The transparency of this process strengthens the trustworthiness of AI models and facilitates regulatory compliance in sensitive applications [13].

The applications of metaheuristic feature selection between 2020 and 2025 have been diverse and impactful. In bioinformatics, these methods have been used to select minimal gene subsets that preserve diagnostic accuracy, enabling cost-effective and interpretable disease prediction. In medical imaging, they reduce the dimensionality of texture or wavelet features extracted from MRI and CT scans, improving the efficiency of classification models. In cybersecurity, adaptive swarm-based algorithms optimize subsets of network traffic attributes for intrusion detection, achieving high detection rates with minimal computational overhead. In industrial engineering, human–social and physics-based algorithms support process optimization and fault detection by identifying the most relevant monitoring variables. These applications highlight the versatility of metaheuristics and their capacity to generalize across domains with varying data structures [14].

Despite these advances, challenges remain. Many metaheuristic algorithms still rely heavily on random initialization, making their performance sensitive to initial conditions. Parameter tuning can be computationally expensive, and convergence behavior can vary significantly across problems. Scalability to extremely high-dimensional data, beyond tens of thousands of features, remains an open issue. Moreover, while hybrid and adaptive algorithms show promise, their increased complexity can make them difficult to analyze theoretically and to reproduce experimentally. Future research is expected to focus on improving stability, reducing energy consumption in computation, and establishing standardized evaluation frameworks to ensure comparability across studies [15].

Looking ahead, several emerging directions define the future of metaheuristic-based feature selection. The first is the integration of self-supervised and transfer learning mechanisms, enabling algorithms to reuse knowledge across similar problems or domains. The second is the incorporation of explainable AI principles, ensuring that optimization outcomes are interpretable and justifiable. The third involves developing lightweight, energy-efficient versions of metaheuristics suitable for real-time and embedded applications. Finally, the coupling of metaheuristics with quantum computing and neuromorphic architectures represents a frontier where optimization can leverage the next generation of computational paradigms.

In summary, feature selection remains a cornerstone of modern data-driven modeling, serving as an essential bridge between raw data and effective learning. The progress achieved from 2020 to 2025 in metaheuristic optimization has redefined how feature subsets are discovered and evaluated. The shift toward hybrid, adaptive, and interpretable frameworks marks the beginning of a new generation of intelligent search system algorithms that do not merely explore data but learn, reason, and adapt within it. Through these developments, metaheuristics are evolving from problem-solving tools into cognitive frameworks that embody the very principles of AI itself.

The remainder of this paper is organized as follows. Section 2 reviews the most relevant works on metaheuristic algorithms and their applications to feature selection, providing a comparative overview of recent trends from 2020 to 2025. Section 3 introduces the prominent families of metaheuristic algorithms, including evolutionary algorithms, physician-inspired algorithms, human and social-based algorithms, and swarm intelligence algorithms, each discussed with representative methods and recent advancements. Section 4 presents a critical discussion and outlines the advantages, limitations, and emerging research directions of these algorithms in feature selection. Finally, Section 5 concludes the paper and highlights future perspectives in developing hybrid and intelligent optimization strategies for high-dimensional data problems.

In recent years, the explosive growth of high-dimensional data has intensified the demand for efficient feature selection and optimization algorithms. From 2020 to 2025, the research landscape has evolved toward integrating machine learning with intelligent optimization methods to enhance performance, interpretability, and scalability. Traditional statistical or deterministic approaches have increasingly proven insufficient for handling complex, nonlinear, and noisy datasets common in biomedical imaging, genomics, and text mining. Consequently, metaheuristic optimization has become a dominant paradigm bridging search efficiency with adaptability across domains [5].

The literature during this period reveals two converging research paths: first, the refinement of feature selection methods that reduce dimensionality while preserving discriminative power; and second, the development of metaheuristic optimization algorithms that balance exploration and exploitation across vast solution spaces. Together, these directions reflect a shift from isolated algorithmic design toward hybrid, adaptive, and interpretable frameworks, a hallmark of post-2020 machine learning research [1].

Post-2020 metaheuristics incorporate chaos and hybridization for feature selection. They start with random solutions, using exploration to explore new areas and exploitation to refine them. Inspirations include nature, physics, human behavior, and swarms. Evolutionary use of mutation-selection with chaos to avoid premature convergence, physics-based model energy-forces. Human-Social integrates memory and social cooperation. Swarm capitalizes on agent interactions to build collective intelligence. This diversity supports feature selection tasks [1].

This section reviews metaheuristic algorithms proposed between 2020 and 2025 for feature selection, organized by algorithmic category. Each method is discussed in terms of its mechanism, representation scheme, and empirical performance. Most algorithms employ binary or Boolean encodings to represent candidate feature subsets, where the optimization process seeks to maximize classification accuracy while minimizing the number of selected features [1], [8].

From 2020 to 2025, metaheuristic algorithms have achieved remarkable progress in addressing the NP-hard nature of feature selection. As data complexity and dimensionality have grown, traditional deterministic methods have struggled to cope with exponential search spaces, leading to a surge in adaptive, stochastic, and hybrid metaheuristics. These algorithms have proven highly effective in selecting compact, relevant feature subsets that enhance classification accuracy, reduce redundancy, and improve interpretability across diverse fields from bioinformatics and imaging to cybersecurity and finance.

Over two hundred metaheuristic algorithms were introduced in these five years, reflecting a vibrant era of innovation. Their development can be broadly categorized into four paradigms: evolutionary, physics-based, human–social, and swarm intelligence. Evolutionary algorithms, which account for about 13% of new methods, focus on adaptive evolution and genetic operators that preserve population diversity and generalization. Physics-based algorithms, about 12%, rely on physical modeling principles, such as motion and energy balance, to ensure stable convergence. Human–social algorithms, accounting for roughly 20%, mimic collective human behaviors such as cooperation, competition, and negotiation, offering dynamic adaptability and interpretability. Swarm intelligence algorithms dominate the field, accounting for approximately 55% of new proposals, reflecting their scalability, distributed architecture, and superior performance in high-dimensional feature selection tasks.

This distribution reveals a broader research trend: a shift toward collective, adaptive, and hybrid intelligence. Swarm algorithms are favored for their balance of exploration and exploitation, while human–social methods capture behavioral complexity suited to modern, data-driven systems. Hybrid frameworks that combine multiple paradigms, such as integrating evolutionary diversity with swarm-based refinement or coupling physics-based stability with adaptive learning, have further enhanced performance. Reinforcement learning and chaotic dynamics have also been incorporated, allowing algorithms to self-tune and maintain exploration diversity. The result is a new generation of intelligent optimizers that not only search but also learn, adapt, and self-regulate.

Empirical studies consistently demonstrate that these modern metaheuristics can reduce feature dimensionality by 60–90% while preserving or improving classification accuracy. Their success has established them as a cornerstone of modern machine learning pipelines, enabling faster training, improved generalization, and greater transparency. Moreover, their ability to operate efficiently in high-dimensional and noisy environments positions them as essential tools for real-world applications where scalability and robustness are critical. Nevertheless, challenges remain. Many algorithms depend on random initialization and require careful parameter tuning to achieve optimal results. Ensuring convergence stability, improving computational efficiency, and enhancing reproducibility remain essential research directions. The increasing complexity of hybrid systems also calls for clearer theoretical foundations and standardized evaluation frameworks.

Looking forward, metaheuristic feature selection is moving toward greater intelligence, explainability, and sustainability. Future designs are expected to emphasize interpretable decision processes, online adaptability, energy-efficient computation, and integration with emerging paradigms such as reinforcement learning, transfer learning, and quantum optimization. In summary, the 2020–2025 period represents a transformative era for metaheuristics in feature selection. Evolutionary algorithms contribute adaptive evolution; physics-based methods ensure balance and precision; human–social strategies introduce behavioral intelligence; and swarm intelligence delivers collective scalability. Together, these paradigms have redefined optimization, transforming feature selection from a computational challenge into an intelligent, evolving process that mirrors the principles of learning and cognition.

The authors declare no conflict of interest.

All data are included in the text.

This research received no specific grant from funding agencies in the public, commercial, or not-for-profit sectors.