Hierarchical representations of network data with optimal distortion bounds
Zane Smith, Samir Chowdhury, Facundo Mémoli · 2016
Single linkage hierarchical clustering is a tool in unsupervised learning which has been fully characterized for finite metric spaces, but not for the unrestricted setting of general networks. We follow a recent line of work to complete the characterization for general networks, and moreover, we provide quantitative bounds on how much information is lost when applying our method to network data. These bounds are novel even in the setting of finite metric spaces. Finally, we propose a construction called a treegram that provides a visual summary of the result of applying our method to a network data set.