Metaclustering

Metaclustering is a data analysis technique in unsupervised learning in which the results of multiple clustering analyses are combined or reorganized to produce higher-level groupings. In practice, metaclustering can mean clustering the clusters produced by different algorithms or parameter settings, effectively clustering the results of several base clusterings. It is related to ensemble clustering or consensus clustering, where robustness and stability are emphasized.

There are two main flavors. In the co-association approach, many clustering runs are executed on the same data under varying parameters or subsamples, and a matrix is built that records how often every pair of data points co-appears in the same cluster. A final clustering is then applied to this matrix to obtain a consensus partition. In the second flavor, each base cluster is treated as an object with descriptive features (such as centroid, size, and dispersion), and these clusters are themselves clustered to form meta-clusters. This can reveal higher-level structure among the clusters and may help interpret heterogeneous base results.

Applications span bioinformatics, imaging, and marketing, including identification of robust gene modules in microarray or RNA-seq data, integrating segmentation results in computer vision, and stabilizing customer segments across methods. Benefits include increased robustness to initialization and method bias, and a potential reduction of spurious small clusters. Drawbacks include determining the number of meta-clusters, sensitivity to the set of base clusterings, and higher computational cost. Evaluation typically relies on internal stability measures, such as the coherence of the co-association matrix or agreement with external references when available, and on the interpretability of the final meta-clusters.

Metaclustering complements other ensemble techniques and provides a framework for deriving stable, interpretable cluster structures from diverse clustering results.