Graph Embedding and Field Based Detection of Non-Local Webs in Large Scale Free Networks

Scalable graph algorithms are usually discrete and heuristics-based and their performance may strongly depend on the topology and size of the targeted graph. In this paper we investigate an alternative approach where we translate graph problems into a continuous space problem: we embed graphs into an higher dimensional Euclidean space using an N-body approach. Then we define a synthetic field as function of the Euclidean space vertex locations and find peaks in the field. These peaks are used to identify regions and thus vertices of interest, and are an emergent property. We test the approach to identify a low-degree web that is connecting otherwise unrelated vertices in a power law graph. We implement the algorithm in CUDA and show good scalability to graphs with 7 million nodes.