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.
CITATION STYLE
Guo, T., & Zhu, X. (2014). Super-graph classification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 8443 LNAI, pp. 323–326). Springer Verlag. https://doi.org/10.1007/978-3-319-06608-0_27
Mendeley helps you to discover research relevant for your work.