Learning Exact Graph Edit Distance

Abstract

In the previous chapter, two different strategies have been proposed for improving the general distance quality of the approximation framework BP-GED. In the present chapter, two additional approaches are pursued for reducing the approximation error. First, in Sect. 5.1 we introduce a method that aims at estimating the exact edit distance $d_{\lambda _{\min }}(g_1,g_2)$ based on both distance bounds $d_{\psi }(g_1,g_2)$ and $d'_{\psi }(g_1,g_2)$ derived from BP-GED. This method is based on regression analysis. Second, in Sect. 5.2 a novel methodology, which is able to predict the incorrect node operations, is introduced. More precisely, a comprehensive set of features—which numerically characterizes individual node edit operations—is defined and extracted from the underlying graphs. These features are in turn used for the development of a classification model for node edit operations.