On-line learning the graph edit distance costs

Abstract

Abstract This paper presents the first on-line learning method to automatically deduce the insertion, deletion and substitution edit costs of the graph edit distance. The learning method is based on embedding the substitution and deletion operations into a Euclidean space. The points in this space are classified into the ones that represent substitution edit operations and the ones that represent deletion edit operations. Thus, the learning strategy is based on deducing the hyper-plane in this space that best splits these two types of points. Any linear classifier can be used to deduce this hyper-plane, for instance LDA or SVM. The on-line method has the advantage that learning the edit costs and computing the graph edit distance with the new updated costs can be done simultaneously. Experimental validation shows that the matching accuracy is competitive with the off-line methods but without the need of the whole learning set.