Towards a Stable Graph Representation Learning Using Connection Subgraphs

Abstract

This chapter studies the problem of learning large-scale graph representations (a.k.a. embeddings) that encode the relationships among distinct nodes. The learned representations generalize over various tasks, such as node classification, link prediction, and recommendation. Learning node representations aims to map proximate nodes close to one another in a low-dimensional vector space. Thus, embedding algorithms pursue to preserve local and global network structure by identifying node neighborhoods. However, many existing algorithms generate embeddings that fail to preserve the network structure and are unstable. That is, the embeddings yield from multiple runs on the same graph are different. Therefore, on one side of the spectrum, such algorithms seem to be suitable for single graph-related tasks, like node classification; however, on the other side, these algorithms cannot fit multi-graph problems.

In this chapter, we propose a novel stable graph representation learning using connection subgraphs (GRCS) algorithmic framework, which aims to learn graph representations using connection subgraphs, where analogy with electrical circuits has been employed. The connection subgraphs are known to be very beneficial in different real-world networks, such as social networks, biological networks, citation networks, co-authorship networks, terrorism networks, and others, as they address the proximity among each two non-adjacent nodes, which are abundant in real-world networks, by maximizing the amount of flow between them. Although a subgraph captures proximity between two non-adjacent nodes, the formation of the subgraph accounts for connections with immediate neighbors as well. In addition, using connection subgraphs, we address the issues of high-degree nodes, and take advantage of weak ties and use the meta-data that have been neglected by embedding baseline algorithms.

We demonstrate the efficacy and robustness of GRCS over existing representation learning algorithms on a node classification task using data sets from various domains. GRCS is robust to noise; its performance is either as good as or better than that of the state-of-the-art algorithms.