Super-graph classification

Abstract

Graphs are popularly used to represent objects with dependency structures, yet all existing graph classification algorithms can only handle simple graphs where each node is a single attribute (or a set of independent attributes). In this paper, we formulate a new super-graph classification task where each node of the super-graph may contain a graph (a single-attribute graph), so a super-graph contains a set of inter-connected graphs. To support super-graph classification, we propose a Weighted Random Walk Kernel (WRWK) which generates a product graph between any two super-graphs, and uses the similarity (kernel value) of two single-attribute graph as the node weight. Then we calculate weighted random walks on the product graph to generate kernel value between two super-graphs as their similarity. Our method enjoys sound theoretical properties, including bounded similarity. Experiments confirm that our method significantly outperforms baseline approaches. © 2014 Springer International Publishing.