Large Weighted Graph Layouts by Deep Learned Multidimensional Projections

Abstract

tsNET is able to create very high quality graph layouts, but is to slow to run on large graphs. We propose a new graph layout method, NNP-NET, based on tsNET, with the aim of generating layouts for very large graphs. NNP-NET uses NNP to approximate the t-SNE step of tsNET with neural networks with a similar quality compared to layouts generated by tsNET. This thesis will go into the challenges of adapting NNP to a graph layout context and how we solved them. NNP-NET is compared to other state of the art methods, were we show that NNP-NET gets good quality results when compared to other fast methods. Here we also show that NNP-NET is able to create layouts for graphs with millions of nodes in a reasonable amount of time. For very large graphs, the execution time of NNP-NET ends up lower than competing state of the art methods.