Application of k-step random walk paths to graph kernel for automatic patent classification

Abstract

In this study, we focus on utilizing the patent citation graph structure. We investigate the effect of using only one document feature which is patent class along citation graph for the classification task. We collect advantages of a kernel-based method and build kernel function to represent feature and citation associated information. We use k-step random walk paths algorithm to calculate kernel values of each patent pairwise and SVM classifier to do the classification task. We employ sub graph technique for a large patent graph to represent citation graph information. The method is based on the property of neighborhood in a graph. The evaluation of the k-step random walk paths kernel metrics on three datasets from the United States Patent and Trademark Office (USPTO) database shows that using patent citation graph structure with only one feature achieved better performance than previous studies.