Co-clustering of nonsmooth graphons (Q2403427)

The author considers co-clustering or co-blockmodels, i.e., the rows and the columns of the data matrix are samples from row and column populations and correspond to the vertices of a bipartite graph. The underlying generating distribution is not specified and may be arbitrarily nonsmooth, that is, we have a nonsmooth graphon model. The author presents an approximation result implying that the blockmodel can detect the structure in the underlying population. In more detail, a bound for the approximation error of the approximation of a blocked version of the generative graphon by the estimated blockmodel is derived. This result means, that ``if a co-clustering `reveals structure' in the data matrix, then similar structure will also exist at the population level.'' Furthermore, extensions of the blockmodel, such as the degree-corrected blockmodels and random dot product graphs are investigated. The population risk is defined to characterize the difference between the actual and the estimated model. In these models error rates depend on the dimensionality of the latent variable space. The presentation is completed by a simulation study.