Learning feature representation for subgraphs

Abstract

Graphs are a rich and versatile data structure. They are widely used in representing data like social networks, chemical compound, protein structures. Analytical tasks against graph data attracted great attention in many domains. Effective graph analytics provides users deep insights of the data. However, due to the structural characteristics of graphs, computation cost for graph analytics tasks on large graph data set can be very high. We discuss two recent frameworks inspired by the advancements in feature representation learning, neural networks and graph kernels, namely patchy-san and subgraph2vec. We conducted experiments with patchy-san and subgraph2vec frameworks for graph classification problems. With established benchmark datasets, we demonstrate that these two frameworks, despite taking different approaches, are efficient and competitive with state-of-the-art techniques.