Applications of deep network signatures to subgraph classification, quantification of network structure and topologically heterogeneous node classification

Abstract

1. sample a large network such that downstream computations on the sparsified network produce results that are faithful to the full network. Given a large network, we employ several ways of obtaining its sparsified version and then carefully reconstructing it. We test their performances with the task of clustering in mind. We show that it is feasible to capture useful information from partially known graphs and that it can be done very efficiently.

4. identify different behaviors occurring in different regions of topologically heterogeneous networks. We build a lens that one can metaphorically hover on a heterogeneous network to discover different behaviors manifested by the network in different areas. We extract local network signatures and classify the nodes and do significantly better than random. We also show that highly structured networks show high homogeneity.

3. study network structure. We introduce a novel concept to quantify network structure. We build on the success of the lossless image embedding feature to show that different networks have different intrinsic scales at which they exhibit structure.

2. extract network signatures and classify networks with high accuracy. We transform the problem of graph classification into one of image classification. We show that using lossless image features, simple machine learning algorithms can classify networks from a wide variety of domains with high accuracy. We show that, compared to graph kernel and classical features based models, deep learning works best.

Large scale networks are omnipresent due to the accumulation of data in today’s world of “big data”. We need new tools and methods to effectively and efficiently make use of the vast information they contain. In this work, we present techniques to