Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets

Available online at www.rajournals.in RA JOURNAL OF APPLIED RESEARCH ISSN: 2394-6709 DOI:10.47191/rajar/v11i11.16 Volume: 11 Issue: 11 November 2025 International Open Access Impact Factor8.553 Page no.- 1080-1090 1080 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets Yesenia Santana-Cardoso1,2, Alicia Martínez-Rebollar1, Eddie Clemente1, Carlos Fernando Moreno-Calderón1, Dante Mújica-Vargas1 1National Technological Institute of Mexico (TecNM) / National Center for Research and Technological Development (CENIDET), Cuernavaca, Morelos, México 2Technological Subsystem University / Technological University of the Northern Region of Guerrero, Iguala Guerrero, México ARTICLE INFO ABSTRACT Published Online: 29 November 2025 Corresponding Author: Alicia Martinez-Rebollar Clustering is a fundamental technique in Data Science used to identify hidden structures and patterns in unlabeled datasets. However, the wide variety of algorithmic families and the heterogeneity of real-world data make the selection of an appropriate clustering algorithm a persistent challenge. This article presents a literature review of clustering algorithms, considering various taxonomies proposed in previous studies and highlighting emerging trends based on dataset meta-features. Our review addresses the main families of clustering algorithms: partitioning, hierarchical, density-based, model-based, and grid-based emphasizing their underlying assumptions, computational properties, and practical advantages. Additionally, comparative studies published over the past two decades are analyzed to identify performance patterns, recurrent limitations, and the scenarios in which each technique tends to perform best. Finally, this article examines contemporary approaches that incorporate meta-features, metalearning strategies, and algorithm recommendation models aimed at reducing the traditional trial-and-error process through systematic mechanisms that link dataset profiles with suitable algorithms. This state-of-the-art review provides an updated and structured overview of the field, highlighting opportunities to advance toward intelligent clustering algorithm selection. KEYWORDS: Clustering Algorithms, Taxonomies, Comparative Studies, Dataset Meta-Features, Unsupervised Learning. I. INTRODUCTION Data clustering is one of the most widely used techniques in Data Science and Machine Learning, as it enables the discovery of internal structure, hidden relationships, and natural groupings within unlabeled datasets [1]. Its applications span numerous domains, including health sciences, engineering, environmental studies, social sciences, marketing, and cybersecurity, where understanding data organization is essential for decision-making, pattern identification, and exploratory analysis [2], [3]. Despite its broad applicability, selecting the most appropriate clustering algorithm remains a considerable challenge due to the diversity of algorithmic families and the inherent heterogeneity of real-world data. Figure 1 shows the conceptual relationship among three fundamental elements: Data Science, machine learning algorithms, and algorithm implementation. At this intersection lie clustering algorithms, which integrate both theoretical foundations and practical considerations for their application in different contexts. Figure 1: Relationship among Data Science, Machine Learning, and Algorithm Implementation. Over the past decades, a wide variety of clustering methods have been proposed, ranging from classical techniques such as k-means and hierarchical clustering to advanced “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1081 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 approaches based on density, probabilistic models, optimization, or grid structures [4]. These algorithms differ not only in their methodological foundations but also in their assumptions, computational behavior, sensitivity to noise, parameter requirements, and suitability for different types of data. As a result, no clustering method consistently outperforms the others across all scenarios. To address this diversity, several taxonomies [5] have been developed that provide conceptual frameworks for organizing clustering algorithms according to their objectives, mechanisms, or structural assumptions. Likewise, multiple comparative studies have evaluated their performance across different datasets, metrics, and scenarios. Although these works contribute valuable insights, their conclusions often depend on the dataset used and rarely generalize to broader contexts. This reinforces the need for systematic approaches that support algorithm selection based on measurable properties of the data. In recent years, research lines based on data metacharacteristics, meta-learning strategies, and algorithm recommendation models have emerged. Meta-characteristics, such as statistical indicators, structural descriptors, geometric properties, and complexity measures, enable the quantitative characterization of datasets and the linkage of these descriptions to the suitability of certain algorithms. These emerging approaches aim to reduce the traditional trial-anderror practice, moving toward more intelligent and datadriven decision-making in unsupervised learning. In this context, the present article provides a structured review of clustering algorithms, with emphasis on their taxonomies, the comparative analyses reported in the literature, and recent approaches that incorporate dataset meta-characteristics. The objective is to identify persistent challenges and highlight opportunities to advance toward a more systematic and intelligent selection of clustering algorithms. This article is organized as follows: Section II presents the theoretical foundations and state of the art; Section III details a comparative analysis of the algorithms examined. Section IV presents the reference datasets; Section V discusses several clustering applications; and finally, Section VI provides the conclusions and future work. II. THEORETICAL FOUNDATIONS AND STATE OF THE ART This section presents the state of the art, synthesizing the most relevant and recent contributions in the specialized literature. It outlines current advances, methodological approaches, and emerging trends that underpin the theoretical framework and delineate the principal research directions associated with the problem under study. A clustering algorithm is an unsupervised machine learning technique that organizes a dataset into groups of elements that are similar to one another. Its application is relevant in contexts where high dimensionality, heterogeneity, and the absence of prior knowledge make it difficult to understand how data are internally structured. Each clustering algorithm has its own advantages and disadvantages, and selecting the appropriate algorithm depends on several factors, such as the type of data, dataset size, data dimensionality, data distribution, result interpretability, and the objectives of the analysis. The main families of clustering algorithms are outlined below, each characterized by distinct methodological foundations and specific approaches to data organization. A. partitional algorithms The k-means algorithm [6] is a well-known clustering technique in the field of unsupervised learning. It consists of partitioning a set of n objects into k ≥ 2 groups, such that the objects belonging to the same group share similar characteristics, while differing from those in other groups. There are various research studies that improve existing clustering methods. An example is the work described in [7], which proposes a new convergence condition consisting of stopping the K-means algorithm when a local optimum is reached or when no further changes of objects between groups occur. In the research described in [8], a criterion is proposed to balance processing time and solution quality in the K-means algorithm. The authors implement an improvement in the algorithm’s convergence phase using the Pareto principle. They evaluate the enhanced algorithm on five Big Data type datasets, significantly reducing processing time and improving solution quality. The K-means algorithm has been widely used in the Big Data domain because large instances are difficult to cluster. The research described in [9] addresses this problem by improving the K-means algorithm for clustering Big Datatype datasets. The authors developed a heuristic called Honeycomb (HC), which is based on the number of dimensions and the number of centroids that form the neighborhood, in order to reduce the number of distance calculations between objects and cluster centers. This approach achieves up to a 90% reduction in processing time and less than a 1% reduction in solution quality. There are different versions and improvements of the KMeans algorithm. In the study presented in [10], an exploratory analysis is described regarding the behavior of the main variants of the K-Means algorithm (Hartigan– Wong, Lloyd, Forgy, and MacQueen) when solving some of the difficult instance sets from the Fundamental Clustering Problems Suite (FCPS), a benchmark reference. These variants are implemented in the R language and allow computing the minimum and maximum intra-cluster distances of the final clustering. Below is the mathematical formulation of the K-Means algorithm. Let N = {𝑥𝑖,...,𝑥𝑛} be the dataset to be partitioned, where 𝑥𝑖∈ℜ𝑑 for i = 1, …, n y 𝑑≥1 is the number of dimensions. “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1082 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 𝑃: 𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝐽(𝑊,𝑀)=∑ 𝑛 𝑖=1 ∑ 𝑘 𝑗=1 𝑤𝑖𝑗𝑑(𝑥𝑖,𝜇𝑗) Subject to ∑𝑘 𝑗=1 𝑤𝑖𝑗 =1,𝑖=1,...,𝑛, 𝑤𝑖𝑗 =0 𝑜 1,𝑖 = 1,...,𝑛,𝑦 𝑗 = 1,...,𝑘, (1) where 𝑤𝑖𝑗 =1 if and only if the object 𝑥𝑖 belongs to cluster 𝜇𝑗 y 𝑑(𝑥𝑖,𝜇𝑗) denotes the Euclidean distance between 𝑥𝑖 y 𝜇𝑗, for 𝑖 = 1,...,𝑛,𝑦 𝑗 = 1,...,𝑘. The K-means algorithm consists of four stages: initialization, classification, centroid computation, and convergence. The Fuzzy C-Means algorithm [11] is also recognized for its data partitioning, similar to K-means. However, its fuzzy partition allows each object to belong to multiple groups with different degrees of membership. There are research studies focused on combining algorithms to improve results. The study presented in [12] proposes a hybrid variant of the Fuzzy C-Means and KMeans algorithms to address large datasets such as those found in Big Data, introducing a new approach called Hybrid OK-Means Fuzzy C-Means (HOFCM), which optimizes the parameter values of the membership matrix. A different improvement of the Fuzzy C-Means algorithm is described in [13]. The authors use the K-Means algorithm to generate the final centroids, which, through a transformation process known as One-Hot encoding, produce the membership matrix required by the Fuzzy C-Means algorithm. To obtain this encoding, the study proposes selecting attributes using entropy weighting and the Box–Cox transformation. This work highlights improvements in solution time, achieving a 5.19% gain in the best case, and a 5.56% improvement in the silhouette coefficient compared to the standard Fuzzy C-Means algorithm. Another application involving heuristics is described in [14], where the authors implement the Particle Swarm Optimization algorithm to select the optimal initial centroids for Fuzzy C-Means. Its performance is evaluated using three metrics Rand Index, F-measure, and Objective Function through which the authors determine that the proposed method converges in less time than the conventional algorithm. Their results highlight an improvement of 0.5% in the objective function value compared to the standard Fuzzy C-Means algorithm. Other methods for improving cluster initialization involve analyzing data density. An example is the study described in [15], in which the authors propose using the Density Peaks Clustering (DPC) algorithm to identify the initial centroids and the groups with which the dataset will be processed. Among the reported results, they describe an accuracy of 84.99% on the dataset known as Jain and an adjusted Rand index value of 48.48%, both of which outperform other implemented algorithms. Below is the mathematical formulation of the Fuzzy CMeans algorithm. Let N = {𝑥𝑖,...,𝑥𝑛} be the dataset to be partitioned, where 𝑥𝑖∈ℜ𝑑 for i = 1, …, n y 𝑐 and is the number of groups where 2≤𝑐≤𝑛. 𝐽𝑚(𝑈,𝑉)=∑ 𝑛 𝑖=1 ∑ 𝑐 𝑗=1 𝜇𝑖𝑗 𝑚||𝑥𝑖−𝑣𝑗||2 2 (2) where 𝑈=𝜇𝑖𝑗 is the degree of membership of each object i to each group j; 𝑉={𝑣1,...,𝑣𝑛} is the set of centroids, where 𝑣𝑗 is the centroid of group j; m is the weighting exponent, which modifies the degree of membership, m > 1 y ||𝑥𝑖− 𝑣𝑗||2 2 denotes the distance calculations from objects to centroids using the Euclidean norm. The optimal clustering of the data X as pairs (𝑈,𝑉) that locally minimize 𝐽𝑚 yields the following expressions. 𝜇𝑖𝑗 =1 ∑𝑐 𝑘=1 (||𝑥𝑖−𝑣𝑗||2 2 ||𝑥𝑖−𝑣𝑘||2 2)1 (𝑚−1); 1≤𝑖≤𝑛; 1≤𝑗≤𝑐 (3) 𝑣𝑗 =∑𝑛 𝑖=1 (𝜇𝑖𝑗)𝑚𝑥𝑖 ∑𝑛 𝑖=1 (𝜇𝑖𝑗)𝑚; 1≤𝑗≤𝑐 (4) where 𝑥𝑖=(𝑥1,𝑥2 ,...,𝑥𝑑) and 𝑣𝑗=(𝑣1,𝑣2 ,...,𝑣𝑑) are vectors that belong to the space ℜ𝑑. B. Hierarchy-Based Algorithms Hierarchical-based algorithms construct groups through successive merging. A popular example within this category is the Agglomerative Herarchical Clustering algorithm [16]. Unlike partitional methods, this algorithm does not require specifying the initial number of clusters; instead, it generates a hierarchical structure called a dendrogram. One application of this algorithm is the integration of data obtained from wireless sensors. The study in [17] describes the application of the Agglomerative Herarchical Clustering algorithm in clustering data extracted from interconnected sensors, addressing the problem of node failures during data transmission within the sensor network. The authors implement this algorithm in two phases: backbone formation and recovery process. In the first phase, the algorithm is applied to determine the cluster centers of the networks, and in the second phase, mobile nodes are deployed at the cluster centers, achieving improved efficiency in data traffic across the nodes. Below is the mathematical formulation of the HAC algorithm. Let N = {𝑥𝑖,...,𝑥𝑛} be the dataset to be partitioned, where 𝑥𝑖∈ℜ𝑑 for i = 1, …, n and each object forms an individual group 𝐶(0) ={𝑥1,𝑥2,...,𝑥𝑛}, the distance between objects is computed 𝐷(𝑖,𝑗)=||𝑥𝑖−𝑥𝑗||2. 𝛥(𝐴,𝐵) “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1083 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 where A and B are two clusters, and the linkage criterion is defined. 𝐷𝑚𝑖𝑛(𝐶𝑖,𝐶𝑗)=𝑚𝑖𝑛𝑥∈𝐶𝑖,𝑦∈𝐶𝑗||𝑥−𝑦||2 (5) 𝐷𝑚𝑎𝑥(𝐶𝑖,𝐶𝑗)=𝑚𝑎𝑥𝑥∈𝐶𝑖,𝑦∈𝐶𝑗||𝑥−𝑦||2 (6) 𝐷𝑎𝑣𝑔(𝐶𝑖,𝐶𝑗)= 1 |𝐶𝑖||𝐶𝑗|∑ 𝑥∈𝐶𝑖∑ 𝑦 ∈𝐶𝑗||𝑥−𝑦||2 (7) 𝐷𝑤𝑎𝑟𝑑(𝐶𝑖,𝐶𝑗)= |𝐶𝑖||𝐶𝑗| |𝐶𝑖|+|𝐶𝑗|||𝑣𝑖−𝑣𝑗||2 2 (8) where 𝑣𝑖 y 𝑣𝑗 are the centroids 𝐶𝑖 and 𝐶𝑗 respectively. C. Density-Based Algorithms Density-based algorithms create clusters by identifying regions in the n-dimensional space where dense groups of points exist. One of the most widely used algorithms in this category is DBSCAN [18], which identifies dense groups of points and separates them from sparse regions; it also detects outliers in the point distribution. Some studies focus on the initial parameter values of the DBSCAN algorithm to improve its results. An example is the work in [19], where this problem is addressed using the KNN algorithm to compute one of DBSCAN’s initial parameters, in addition to detecting clusters with heterogeneous density. This enables the proposed method to be more efficient and accurate. Let X = {𝑥𝑖,...,𝑥𝑛} be the dataset to be partitioned, where 𝑥𝑖∈ℜ𝑑, 𝜀>0 is the neighborhood radius and MinPts is the number of points required for a region to be considered dense. The neighborhood of a point is defined in expression 9. 𝑁𝜀(𝑥𝑖)={𝑥𝑗∈𝑋|𝑑(𝑥𝑖,𝑥𝑗)≤𝜀} (9) where 𝑑(𝑥𝑖,𝑥𝑗) is the Euclidean distance between the points 𝑥𝑖 y 𝑥𝑗. A point may be considered a core point or a border point; otherwise, it is classified as noise. To determine this, expressions 10 and 11 are evaluated for each point. |𝑁𝜀(𝑥𝑖)|≥ 𝑀𝑖𝑛𝑃𝑡𝑠 (10) |𝑁𝜀(𝑥𝑖)|< 𝑀𝑖𝑛𝑃𝑡𝑠 y ∃𝑥𝑗 núcleo 𝑥𝑖∈𝑁𝜀(𝑥𝑗) (11) D. Model-Based Algorithms Model-based clustering assumes that data are generated from an underlying probabilistic model, typically formulated as a finite mixture of parametric distributions. Each cluster corresponds to one component of the mixture, which enables flexible representation of complex and overlapping structures. These methods provide probabilistic membership assignments and are widely applied in bioinformatics, pattern recognition, and statistical learning due to their ability to incorporate uncertainty and prior knowledge [20], [21]. Given a dataset X = {𝑥𝑖,...,𝑥𝑛} and K mixture components, the model is defined as: (12) where πk represents the mixing coefficients such that is the probability density of component k, and denotes all model parameters. Typically using the Expectation–Maximization (EM) algorithm, which iteratively updates posterior responsibilities and parameter estimates until convergence [21]. E. Grid-Based Algorithms Grid-based clustering algorithms partition the data space into a finite number of non-overlapping cells or grid units and perform clustering at the grid level rather than directly on individual data points. This approach significantly reduces computational complexity, as the clustering process depends on the number of grid cells instead of the number of instances. These methods are highly efficient, scalable, and robust to noise, making them suitable for large datasets or highdimensional applications such as spatial data mining, image analysis, and multidimensional pattern detection [22], [23]. Given a dataset X = {𝑥𝑖,...,𝑥𝑛}, the data space is divided into a grid of cells, and density measures are computed for each cell to determine whether it belongs to a cluster. Gridbased clustering algorithms generally follow four steps: partitioning the space, calculating cell densities, identifying dense regions, and forming clusters by connecting adjacent dense cells. III. COMPARATION AMONG THE MAIN FAMILIES OF CLUSTERING ALGORITHMS Clustering algorithms can be classified according to multiple taxonomies, given the wide diversity of mathematical principles, structural assumptions, and optimization mechanisms that underpin these algorithms [24]. In general, the algorithms can be grouped into families that share common methodological foundations and similar strategies for clustering datasets. Table 1 presents the families of clustering algorithms widely used by the scientific community. A. Partitional Algorithms Partitional clustering algorithms are characterized by generating a single partition of data into K groups. Their operation is based on the optimization of objective functions that seek to minimize dissimilarity and maximize separation among groups [25]. These objective functions may be associated with variance, Euclidean distance, or fuzzy membership measures. Mathematically, they operate under assumptions of convexity, isotropy, and density homogeneity, which implies that the expected groups typically exhibit spherical or simple ellipsoidal geometry [26]. Most of these techniques rely on iterative processes of reassignment and updating, enabling convergence to stable solutions due to their initialization-dependent nature [27]. “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1084 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 This family has been widely used in industrial and scientific applications owing to its efficiency, simplicity, and interpretability [28], particularly in contexts with wellseparated structures or low dimensionality. B. Hierarchy-Based Algorithms Algorithms belonging to this class establish multilevel relationships among instances through a structure known as a dendrogram, which describes the relative similarity between elements in the dataset. Unlike partition-based clustering algorithms, they do not require specifying the number of groups and allow the analysis of clustering structure at multiple scales [16]. Their operation is based on two strategies: agglomerative, which begins with each data point as an individual group; and divisive, which starts with a global group and successively divides it. The key operational criterion is the linkage function, whose selection directly affects the shape of the dendrogram and the final structure of the groups. Mathematically, hierarchical algorithms work with distance matrices, providing an appropriate representation for clustering, though computationally costly [29]. These methods are useful in exploratory analysis, bioinformatics, and social sciences, where interpretability is a critical factor [17]. C. Density-Based Algorithms Clustering algorithms can be classified according to data density; that is, groups are defined as regions in the space where data concentration is higher compared to surrounding areas [30]. This formulation derives from notions of topology, graph theory, and density connectivity. Generally, these methods identify dense connected components within a geometric graph constructed using metric distances. They are suitable for non-convex structures and datasets containing noise. These methods assume that a cluster is a densely connected region in which each point has a minimum number of neighbors within a given radius; however, high sensitivity to global parameters may lead to unstable results. Recent studies report their effectiveness in spatial and geographic analysis for detecting dense regions in GPS data and urban structures [18], as well as in contemporary anomaly-detection tasks applied to network traffic, cybersecurity, and sensor monitoring in IoT environments [31], [32]. D. Grid-Based Algorithms This type of clustering algorithm quantifies the data space into a finite number of cells or grids, within which local statistics are computed to represent the regional distribution of the data. This approach reduces computational complexity, as the dataset is scanned only once to construct the grid, while subsequent clustering occurs in a transformed space whose size depends on the selected resolution and the total number of instances. Notable advantages arise in terms of efficiency, scalability, and noise resistance, making these methods suitable for large or high-dimensional databases. The quality of the clustering depends on the size of the cells into which the data space is divided. The literature reports its application in spatial data mining, geographic analysis, and the processing of information structured in subspaces [33], [34]. E. Model-Based Algorithms This type of algorithm formulates the clustering problem as a parameter estimation task, where each cluster is interpreted as a component of a mixture distribution; the dataset is modelled as a weighted combination of parametric distributions, allowing the underlying statistical structure of the groups present in the data to be represented [35]. A distinguishing characteristic of this family is the geometric flexibility derived from the parameterization of covariance matrices, which may assume spherical, diagonal, elliptical, or freely oriented forms [36]. Recent research highlights the growth of Bayesian models, particularly Bayesian Gaussian Mixtures, which are used in scenarios with limited labelled data, complex structures, and the need to incorporate prior knowledge [37]. The broad and growing variety of clustering algorithms characterized by diverse theoretical foundations, statistical assumptions, and operational mechanisms has turned the selection of the most appropriate algorithm for a specific problem into a methodological challenge. In this context, the construction of a formal taxonomy of clustering algorithms emerges as a fundamental tool for structuring existing knowledge. In this way, classifying algorithms into coherent conceptual families can establish a reference framework that facilitates their systematic comparison and justified application in particular scenarios [38]. Table 1 presents the algorithms that make up the partitional family, including methods widely used for their efficiency and suitability for exploratory analysis in large datasets. Table 2 includes the algorithms belonging to the hierarchical family, recognized for constructing nested relationships among data through divisive or agglomerative processes. Table 1. Taxonomy of partition-based clustering algorithms. Algorithm Metrics Used Advantages Disadvantages Usage Scenarios Ideal Dataset Characteristics K-means Euclidean, Cosine Fast and efficient; easy to implement. Sensitive to outliers; requires k; convex clusters. Numerical data, rapid segmentation. Numerical variables; spherical clusters; no outliers. “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1085 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 K-medoids Euclidean, Manhattan, Hamming Robust to outliers; supports multiple metrics. More computationally expensive. Mixed data. Mixed variables; presence of outliers. PAM Euclidean, Manhattan, Hamming Highly robust; high quality for small datasets. High computational cost. Qualitative or specialized analysis. High precision; small and noisy datasets. CLARA Euclidean, Manhattan, Hamming More scalable than PAM. Depends on sampling. Large-scale data mining. Representative samples; homogeneous distribution. CLARANS Euclidean, Manhattan, Hamming Balance between speed and accuracy. Increased randomness. Semi-automated clustering. Complex data; interactive exploration. MiniBatch Kmeans Euclidean Very fast; scalable; low memory usage. Lower accuracy. Streaming, big data, online recommendations. Large volumes; stable streaming data. Algorithm Metrics Used Advantages. Disadvantages. Usage Scenarios. Ideal Dataset Characteristics. Table 2. Taxonomy of hierarchy-based clustering algorithms. Algorithm Metrics Used Advantages Disadvantages Usage Scenarios Ideal Dataset Characteristics Agglomerative Hierarchical Linkage distance Does not require k initially; dendrogram visualization. High computational cost; sensitive to outliers. Structural exploration; biological or linguistic studies. Numerical data with clear hierarchical structure. Divisive Hierarchical Linkage distance Explores large hierarchical structures. Less supported in libraries; results sensitive to early decisions. Community analysis or hierarchical classification. Nested, nonspherical structures. Ward’s Method Intra-cluster variance; Euclidean distance Reduces variance; produces compact and balanced clusters. Assumes spherical shape; not suitable for elongated clusters. Clustering in psychometrics and marketing studies. Homogeneous data; expected compact clusters. Table 3 groups the algorithms corresponding to the densitybased family, characterized by identifying regions of high data concentration. Table 4 covers the algorithms that form part of the model-based family, oriented toward representing data through statistical distributions. Table 5 compiles the algorithms that constitute the grid-based family, which partition the space into discrete cells to operate on local aggregations. Table 3. Taxonomy of density-based clustering algorithms. Algorithm Metrics Used Advantages Disadvantages Usage Scenarios Ideal Dataset Characteristics DBSCAN Euclidean distance (customizable) Identifies arbitrarily shaped clusters; detects noise. Sensitive to parameters; difficulty in highdimensional spaces. Spatial databases, anomaly detection. Variable density; nonspherical clusters; presence of noise. OPTICS Euclidean distance or user-defined metric Does not require specifying the number of clusters; handles variable densities. High computational complexity; requires interpretation of the Reachability plot. Data with density variability; detailed exploration. Overlapping clusters; heterogeneous densities. “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1086 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 HDBSCAN Euclidean distances; density-based metrics Improved detection of complex structures and noise; natural hierarchy. Requires finetuning; results may be less intuitive. Data science, bioinformatics, geospatial analysis. Large volumes; hierarchical density structure. Table 4. Taxonomy of model-based clustering algorithms. Algorithm Metrics Used Advantages Disadvantages Usage Scenarios Ideal Dataset Characteristics Gaussian Mixture Models (GMM) Loglikelihood probability, Mahala Nobis distance Allows elliptical cluster shapes; probabilistic. Sensitive to initialization; requires predefined k Probabilistic classification, NLP, image segmentation. Continuous data with overlapping clusters. Hidden Markov Models (HMM) Transition and emission probabilities Models temporal sequences; detects patterns. High training complexity; requires ordered sequences. Speech recognition, time series analysis. Temporal or sequential data. Bayesian Gaussian Mixture Posterior distribution, Bayesian criteria Prevents overfitting; incorporates uncertainty. Higher computational cost; statistical complexity. Problems requiring prior knowledge integration. Complex structures and limited labeled information. Table 5. Taxonomy of grid-based clustering algorithms. Algorithm Metrics Used Advantages Disadvantages Usage Scenarios Ideal Dataset Characteristics STING (Statistical Information Grid) Cell density; aggregated statistics Fast and scalable; low computational cost Fixed resolution; loses precision in complex structures Spatial data mining; range queries Geographic or spatially structured data CLIQUE (Clustering in Quest) Density in subspaces; number of points per cell Identifies clusters in highdimensional subspaces Sensitive to cell size; not always intuitive Multidimensional databases; bioinformatics High-dimensional data with hidden clusters Wave Cluster Frequency energy; implicit spatial distance Captures arbitrary shapes; robust to noise Complex to implement; requires specialized parameters Image processing, signal analysis, spatial pattern detection Spatial data with noise or fractal structure IV. REFERENCE DATASETS The comparison and analysis of clustering algorithms require standardized datasets that allow their behavior to be studied under controlled conditions. In the literature, repositories such as the UCI Machine Learning Repository [39], OpenML [40], and KEEL [41] are relevant due to the variety of datasets they provide. A. UCI Machine Learning Repository The UCI Machine Learning Repository is one of the most influential and widely used repositories by the scientific community since its creation in 1987 [39]. Its original purpose was to provide an accessible collection of datasets for researchers interested in the empirical analysis of machine learning and statistical algorithms. Over time, it has become a reference resource for the development, testing, and comparison of both supervised and unsupervised techniques, including clustering algorithms. The UCI repository provides the scientific community with a wide variety of datasets organized according to different criteria, such as task type, application domain, and attribute characteristics. This structure makes it easier for researchers to select appropriate datasets for evaluating and comparing clustering algorithms based on the specific needs of their experiments. “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1087 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 Among the most representative datasets are those listed in Table 6, which have been widely used in the literature due to their public availability, the quality of their documentation, their consistent presence in comparative studies, and their historical relevance as benchmarks for evaluating clustering algorithms. Table 6: Most frequently used datasets. Datasets Instances Attributes Description Iris 150 4 Flower morphology; classic dataset for clustering validation. Wine 178 13 Chemical analysis of wines; widely used in partitional clustering. Breast Cancer Wisconsin 569 30 Cancer diagnosis; used in clustering studies and external metrics. Seeds 210 7 Geometric characteristics of grains; well-defined groups. MNIST 70,000 784 Image dataset; widely used in clustering, dimensionality reduction, and advanced methods. B. OpenML The OpenML data repository [40] is a collaborative platform designed to facilitate reproducible experimentation in machine learning. Unlike traditional repositories, OpenML not only stores datasets but also integrates metadata, structured descriptions, predefined tasks, workflows, evaluations, and experimental results generated by the scientific community. This architecture, oriented toward traceability and standardization, has made OpenML a key resource for research requiring consistent databases and systematic experimentation in the field of clustering. One of the distinctive features of OpenML is its ability to provide standardized versions of datasets, which allows clustering techniques to be compared under controlled and replicable conditions. The platform maintains a detailed record of each dataset’s properties, including the number of instances, attributes, variable types, complexity measures, basic statistics, and automatically computed meta-features. This approach is particularly relevant for research involving meta-learning, algorithm recommendation models, and analyses based on meta-characteristics. B. KEEL The KEEL repository [41] offers an extensive collection of databases specifically oriented toward machine learning, with a particular emphasis on unsupervised problems, evolutionary methods, and hybrid techniques. Unlike traditional repositories, KEEL includes preprocessed, normalized, and partitioned versions of each dataset, which facilitates the execution of rigorous and fully reproducible comparative analyses. Among its most notable contributions, KEEL provides: • Datasets designed for clustering, feature synthesis, and dimensionality reduction. • Standardized versions of each dataset under different preprocessing levels (normalization, discretization, noise removal), allowing the evaluation of algorithm behavior under controlled scenarios. • Detailed metadata, including statistical information, number of classes (when applicable), complexity levels, presence of noise, and attribute structure. • Special collections, such as hard datasets used for benchmarking evolutionary algorithms, density-based methods, or hierarchical clustering techniques. The UCI, OpenML, and KEEL repositories constitute the main sources for evaluating clustering algorithms, each providing complementary contributions. UCI offers classic and well-documented datasets; OpenML provides standardized versions with complete metadata; and KEEL integrates preprocessed datasets for rigorous comparative analyses. This diversity facilitates reproducible experiments and strengthens the systematic study of algorithmic performance, while also supporting emerging approaches based on meta-characteristics and meta-learning aimed at intelligent algorithm selection. V. CLUSTERING APPLICATIONS Clustering algorithms have a wide range of applications across various domains, where pattern identification, element segmentation, and structural data organization are essential for understanding complex phenomena. In the health domain, clustering has been used to identify clinical profiles, segment patients according to medical characteristics, and characterize disease patterns, supporting diagnosis and therapeutic personalization [42], [43]. In the field of education, clustering techniques enable the creation of learning profiles, the detection of interaction patterns in virtual environments, the classification of academic trajectories, and the analysis of student behaviors, contributing to the development of recommendation systems and learning analytics [44]. In finance and economics, clustering is used to segment customers, identify risk profiles, detect anomalous behaviors, and analyze patterns in economic transactions, strengthening data-driven decisionmaking and financial risk management [45]. In the domain of social network analysis, clustering has proven useful for community detection, grouping users with similar behaviors, and identifying emerging topics, enabling the understanding of social dynamics and global information “Clustering algorithms: taxonomies, comparative studies, and emerging approaches based on meta-features of datasets” 1088 Yesenia Santana-Cardoso1, RAJAR Volume 11 Issue 11 November 2025 trends [46]. Finally, in the field of security, these techniques play a key role in detecting anomalous patterns, identifying suspicious activities, analyzing network traffic, and strengthening behavior-based intrusion detection systems [47]. VI. CONCLUSIONS AND FUTURE WORK The state of the art in clustering algorithms reveals a broad and diverse field, structured into five main families: partitional, hierarchical, density-based, model-based, and grid-based. Each family presents its own characteristics and accepted data types; this methodological diversity confirms the absence of a universally superior algorithm and underscores the importance of considering dataset characteristics in the selection process. The UCI Machine Learning, OpenML, and KEEL repositories, together with the datasets widely used by the scientific community such as: Wine, Breast Cancer Wisconsin, and MNIST, constitute the most frequently used experimental basis in the literature. Their public availability, clear documentation, and ease of use provide methodological consistency, reproducibility, and comparability across studies. Clustering applications in domains such as health, finance, economics, social networks, security, among others, demonstrate its presence across multiple scientific and technological areas. The ability to group information and support data-driven decision-making reinforces its importance in Data Science. The combination of established taxonomies, standardized datasets, and a wide range of applications provides a solid foundation for future developments and continuous improvement of clustering techniques within Data Science. In this context, future work is oriented toward the development of more precise taxonomies, large-scale comparative studies, and approaches based on dataset metacharacteristics that enable the selection of clustering algorithms. Likewise, growth is anticipated in hybrid methods and advanced probabilistic approaches, along with recommendation systems that automate the choice of the most suitable algorithm according to the structure and complexity of the dataset. These lines of research represent a promising trajectory for the use of clustering techniques and their integration into increasingly complex scenarios within Data Science. ACKNOWLEDGMENT The Student Yesenia Santana Cardoso Acknowledges his scholarship (grantee No. 4007298) To the Secretaría de Ciencia, Humanidades, Tecnología e Innovación. This research was partially funded by the National Technology of Mexico with the project: 21639.25-P. REFERENCES 1. Liu, Q., Liu, J., Li, M., & Zhou, Y. (2021). Approximation algorithms for fuzzy C-means problem based on seeding method. Theoretical Computer Science, 885, 146-158. 2. Aalst, W. (2016). Process Mining. En Data Science in Action (págs. 3-23). Springer. 3. Provost, F., & Fawcett, T. (2013). Data Science and Its Relationship to Big Data and Data-Driven Decision Making. Big data, 1(1), 51–59. 4. Pitafi , S., Anwa, T., & Sharif, Z. (2023). A Taxonomy of Machine Learning Clustering Algorithms, Challenges, and Future Realms. Applied Sciences, 13, 3529-3529. 5. Dinh, T., Wong, H., Fournier-Viger, P., Lisik, D., Ha, M.-Q., Hieu Chi, D., & Van-Nam, H. (2025). Categorical data clustering: 25 years beyond Kmodes. Expert Systems with Applications, 272, 249. 6. MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations. Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probability, 1, 281-297. 7. Pérez, J., Montes, P., Mexicano Santoyo, A., & Rodriguez-Jorge, R. (2017). Acceleration of the Kmeans algorithm by removing stable items. International Journal of Space-Based and Situated Computing, 7, 72-81. 8. Pérez Ortega, J., Almanza-Ortega, N., & Romero, D. (2018). Balancing effort and benefit of K-Means clustering algorithms in Big Data realms. PLOS One, 13(9), 1-19. 9. Pérez-Ortega, J., Hidalgo-Reyes, M., CastroSánchez, N., Pazos-Rangel, R., Díaz-Parra, O., Olivares-Peregrino, V., & Almanza-Ortega, N. (2018). An Efficient Heuristic Applied to the Kmeans Algorithm for the Clustering of Large Highly-Grouped Instances. Computing and Systems, 2. 10. Almanza-Ortega, N., Pérez-Ortega, J., Zavala-Díaz, J., & Solís-Romero, J. (2022). Comparative Analysis of K-Means Variants Implemented in R. Computing and Systems, 26(1). 11. Bezdek, J., Ehrlich, R., & Full, W. (1984). FCM: The Fuzzy c-Means clustering algorithm. Computers & Geosciences, 10(2), 191-203. 12. Pérez Ortega, J., Roblero Aguilar, S., AlmanzaOrtega, N., & Frausto, J. (2022). Hybrid Fuzzy CMeans Clustering Algorithm Oriented to Big Data Realms. Axioms, 11, 377. 13. Wu, Z., Yao, J., & Chen, G. (2019). The Stock Classification Based on Entropy Weight Method