Title: Density-peak clustering of graphs
Authors: Riccardo Giubilei - Luiss Guido Carli (Italy) [presenting]
Abstract: Graph clustering, intended as the task of grouping observations that are in the form of graphs, is attracting increasing attention thanks to its various applications. These include identifying similar brain networks for ability assessment or disease prevention, as well as clustering different snapshots of the same network evolving over time to identify similar patterns or abrupt changes. However, there are no well-established procedures for performing this task. The method proposed here builds upon the density-peak algorithm (DP), which is a mode-based clustering approach that identifies cluster centers as data points being surrounded by neighbors with lower density and far away from points with higher density. The new method: 1) inherits the favorable properties of the DP; 2) overcomes two main limitations of the DP, namely, the unstable density estimation and the absence of an automatic procedure for selecting cluster centers; 3) can be applied to graphs of any type, provided that a sensible distance between observations is selected. Numeric applications, including an empirical analysis whose goal is clustering brain connectomes to distinguish between patients affected by schizophrenia and healthy controls, show the adequate performance of the proposed approach.