Universal Graph Embedding Fine Tuning with Dirichlet Energy Smoothing

Abstract

Traditionally, networks are interpreted as discrete entities with nodes connected by links. In this work we propose to interpret networks as fields describing the distribution of certain properties in the multidimensional space. By following the field interpretation of networks, we introduce a universal fine-tuning of node embed dings using the concept of Dirichlet energy smoothing to obtain desirable properties of node embeddings.