Clustering Techniques in High-Dimensional Data

Clustering in high-dimensional spaces presents unique challenges arising from the so-called “curse of dimensionality”, where the volume of the feature space grows exponentially and distances between points become less informative. Conventional algorithms such as k-means and hierarchical clustering frequently suffer from degraded performance due to noise, sparsity and the need to predefine the number of clusters. In response, a broad spectrum of specialised techniques has been developed. Subspace and projected clustering methods seek to identify relevant feature subsets in which cluster structure is more pronounced, thereby reducing noise dimensions. Spectral clustering techniques exploit eigenstructures of similarity graphs to capture non-linear separations, while density-based algorithms have been adapted to high-dimensional settings by employing adaptive neighbourhood definitions. Manifold learning approaches and graph-based embeddings enable the discovery of intrinsic low-dimensional structures prior to clustering. Multi-view and co-clustering methods integrate complementary information from heterogeneous feature groups. Recent advances also include self-expressive models that represent each point as a combination of others, with optimisation strategies tailored to large-scale data through parallel and distributed frameworks. Dimensionality reduction—via principal component analysis, manifold approximation or deep autoencoders—remains a pivotal preprocessing step, often combined with feature selection to maintain interpretability. Together, these innovations have broadened the applicability of clustering to domains such as single-cell genomics, hyperspectral imaging, text mining and recommender systems, where extracting meaningful groupings from massive, noisy feature sets is critical for downstream analyses and decision-making.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent work has introduced a parallelisable multi-subset self-expressive model for subspace clustering that alleviates the computational burden of handling all pairwise relationships. By dividing data into small subsets and solving decomposed optimisation tasks in parallel, this approach preserves global self-expressiveness while achieving substantial speed-ups on synthetic and real-world high-dimensional benchmarks. In another strand of research, a systematic investigation of statistical power for clustering pipelines has revealed that clear subgroup separation is essential for reliable partitioning, especially when dimensionality reduction techniques such as multi-dimensional scaling are applied. This study also demonstrated that fuzzy clustering and mixture modelling provide more robust classification of overlapping multivariate patterns than hard-partition methods, guiding practitioners on sample size requirements and algorithm choice. Furthermore, a comprehensive review of feature dimensionality reduction methods has compared feature selection and extraction strategies, assessed their effectiveness on small-sample and deep-learning scenarios, and highlighted the trade-off between information preservation and computational efficiency. This review underscores the importance of integrating dimensionality reduction with clustering to mitigate noise and enhance interpretability in complex, large-scale datasets.

Clustering Techniques in High-Dimensional Data publication trend

The graph below shows the total number of articles in clustering techniques in high-dimensional data across all publications each year (not limited to Nature Index journals).

Technical terms

High-dimensional data: Datasets in which the number of features greatly exceeds the number of observations, leading to sparsity and diminished discriminatory power of distance metrics.

Curse of dimensionality: A phenomenon where increasing dimensionality causes exponential growth of the feature space, making statistical estimation and clustering less reliable.

Subspace clustering: A family of methods that identify clusters within subsets of features, thereby isolating informative dimensions and reducing noise.

Self-expressive model: An approach where each data point is represented as a linear combination of other points, enforcing a sparse or structured representation that reveals subspace membership.

Dimensionality reduction: Techniques—such as principal component analysis, manifold embedding or autoencoders—that project high-dimensional data onto lower-dimensional representations while retaining essential structure.

Fuzzy clustering: A clustering paradigm that assigns degrees of membership to clusters for each data point, accommodating overlapping group structures.

References

1. PMSSC: Parallelizable multi-subset based self-expressive model for subspace clustering. Computational Visual Media (2023).
2. Statistical power for cluster analysis. BMC Bioinformatics (2022).
3. Feature dimensionality reduction: a review. Complex & Intelligent Systems (2022).