A Topological Approach to Spectral Clustering
This addresses clustering problems in data analysis by offering a method that avoids specifying cluster counts, though it appears incremental as it builds on spectral clustering with topological criteria.
The paper tackles unsupervised clustering by proposing two algorithms that use topological models to identify connected components from data sampled uniformly on a metric space, outputting clusters without requiring prior knowledge like the number of clusters.
We propose two related unsupervised clustering algorithms which, for input, take data assumed to be sampled from a uniform distribution supported on a metric space $X$, and output a clustering of the data based on the selection of a topological model for the connected components of $X$. Both algorithms work by selecting a graph on the samples from a natural one-parameter family of graphs, using a geometric criterion in the first case and an information theoretic criterion in the second. The estimated connected components of $X$ are identified with the kernel of the associated graph Laplacian, which allows the algorithm to work without requiring the number of expected clusters or other auxiliary data as input.