Evolving Beyond Gaussian Prototypes with Multi-Scale Graph Consensus on Microclusters
Year: 2026
Type: conference-paper
Abstract: In this paper, we propose Evolving Microcluster Graph Consensus (EvoGC), a clustering method that combines online, single-pass data compression with offline graph-based cluster recovery. Existing evolving approaches must discover the number of clusters and assign labels in a single pass over the data, placing a heavy burden on a mechanism that processes each sample only once. EvoGC decouples these tasks. The evolving component reduces the input into a fine-grained set of microclusters, deliberately oversegmenting the data, with a support radius derived automatically from intrinsic dimensionality. The microclusters are not interpreted as the final clusters but serve as local building blocks from which clusters of arbitrary shape and scale are described. In the second stage, shared-nearest-neighbour graphs are constructed at multiple neighbourhood scales, partitioned through density-peak-seeded graph merging, and combined using spectral co-association consensus. Evaluation on 17 standard clustering benchmarks spanning manifold, Gaussian, and high-dimensional structures shows that EvoGC achieves the highest average accuracy and lowest average rank compared to nine established methods, with a fixed parameter configuration and no per-dataset tuning.
Authors: Miha Ožbot, Igor Škrjanc
Institution: University of Ljubljana
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225758328
