Analyzing Centralities of Embedded Nodes

Abstract

Given a dataset described as a graph such as social networks, node embedding algorithms estimate a real-valued vector for each node that can later be used for a machine learning task such as node classification. These embedding vectors simplify the task and often improve the task performance. Although word embeddings, e.g., skip-gram and CBOW, have been well analyzed, little is known about the properties of node embeddings. In this paper, we analyze empirical distributions of several node centrality measures, such as PageRank, based on node classification results. Experimental results give insights into the properties of embeddings, which can provide cues to improve embedding algorithms.