Abstract
The success of graph embeddings or nodrepresentation learning in a variety of downstream tasks, such as node classification, link prediction, and recommendation systems, has led to their popularity in recent years. Representation learning algorithms aim to preserve local and global network structure by identifying node neighborhoods. However, many existing network representation learning methods generate embeddings that are still not effective enough, or lead to unstable representations due to random processes (e.g., random walks to generate context) and thus, cannot generalize to multi-graph problems. In this paper, we propose SURREAL, a novel, stable graph embedding algorithmic framework that leverages “spatio-electric” (SE) subgraphs: it learns graph representations using the analogy of graphs with electrical circuits. It preserves both local and global connectivity patterns, and addresses the issue of high-degree nodes that may incidentally connect a pair of nodes in a graph. Further, it exploits the strength of weak ties and meta-data that have been neglected by baselines. The experiments show that SURREAL outperforms state-of-the-art techniques by up to 37% (6% on average) on different multi-label classification problems. Further, in contrast to baseline methods, SURREAL, being deterministic, is stable and thus can generalize to single and multi-graph tasks.
Highlights
Conventional graph mining algorithms (Goyal and Ferrara 2017) have been designed to learn a set of hand-crafted features that best perform to conduct a specific downstream task; i.e., link prediction (Liben-Nowell and Kleinberg 2007), node classification (Bhagat et al 2011), and recommendation (Yu et al 2014)
Recent research efforts have focused on designing either unsupervised or semi-supervised algorithms to learn node representations (Perozzi et al 2014; Perozzi et al 2016; Grover and Leskovec 2016; Tang et al 2015). Such efforts have been initiated in the domain of natural language processing (NLP) (Mikolov et al 2013; Le and Mikolov 2014; Mikolov et al 2013), where two word2vec (Mikolov et al 2013) models have been proposed, namely continuous bag of words (CBOW) and SkipGram
We propose a graph embedding approach that robustly preserves local and global structure by leveraging the notion of network flow to produce approximate but high-quality SE subgraphs between pairs of non-adjacent nodes in undirected andweighted large-scale graphs
Summary
Conventional graph mining algorithms (Goyal and Ferrara 2017) have been designed to learn a set of hand-crafted features that best perform to conduct a specific downstream task; i.e., link prediction (Liben-Nowell and Kleinberg 2007), node classification (Bhagat et al 2011), and recommendation (Yu et al 2014). A unified set of features that can effectively generalize over distinct graph mining-related tasks is exploited To this end, recent research efforts have focused on designing either unsupervised or semi-supervised algorithms to learn node representations (Perozzi et al 2014; Perozzi et al 2016; Grover and Leskovec 2016; Tang et al 2015). Since real-world networks convey more complex relationships compared to those emerging in corpora, some recent representation learning algorithms (Perozzi et al 2014; Perozzi et al 2016; Grover and Leskovec 2016) generate representations that are still not effective enough in preserving network structure, and have room for improvement This in turn impacts the quality of node representations, which compromises the performance of downstream processes. While baseline representation learning methods strive to preserve similarities among nodes in a single graph, they fail to maintain similarities across different runs of the methods, even with using the same dataset (Heimann et al 2018) (graph similarity (Koutra et al 2013) and network alignment (Bayati et al 2009))
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have