B0868
Title: Bayesian nonparametric projected normal mixture models for spectral graph clustering with degree heterogeneity
Authors: Francesco Sanna Passino - Imperial College London (United Kingdom) [presenting]
Abstract: Real-world networks commonly exhibit within-group degree heterogeneity. For example, in an enterprise computer network, users from the same department might have very different levels of activity depending on their job. The objective of this project is to improve existing methodologies for clustering graphs with within-group degree heterogeneity, under the degree-corrected stochastic blockmodel (DCSBM) framework. In previous work, it has been shown that, under the DCSBM, the performance of community detection algorithms based on spectral embedding could be improved by a transformation to spherical coordinates of a scaled spectral decomposition of the graph adjacency matrix, called spectral embedding. A Bayesian nonparametric mixture of projected normals is proposed to perform clustering of nodes on the unit d-sphere resulting from the transformation. The methodology is demonstrated to outperform existing techniques and applied to real data from a university computer network.
