The paper presents a method to learn the substitution, deletion and insertion costs on nodes and edges applied to the graph edit distance. We model the learning strategy as a general model and then we concretise it in two different architectures: the first architecture is based on a neural network and the second architecture is based on a multivariate normal distribution, which have been previously trained. The insertion, deletion and substitution costs on nodes and edges are defined as functions that depend on the output of the machine learning architecture. Other machine learning architectures have been presented in the literature to define the graph edit distance costs. Nevertheless, the main feature of our method is that the insertion, deletion and substitution costs are learned together, training the same machine learning and generating only one model. Thus, these costs are influenced one to another, achieving a higher accuracy in the pattern recognition stage than previous methods.

本文提出了一种学习应用于图编辑距离的节点和边上的替换，删除和插入成本的方法。我们将学习策略建模为一个通用模型，然后将其具体化为两种不同的体系结构：第一种体系结构是基于神经网络的，第二种体系结构是基于多元正态分布的，先前已经对其进行了训练。节点和边缘上的插入，删除和替换成本被定义为取决于机器学习架构输出的函数。其他机器学习架构已在文献中提出，以定义图形编辑距离成本。不过，我们方法的主要特点是，插入，删除和替换成本是一起学习的，训练相同的机器学习并仅生成一个模型。因此，这些成本彼此影响，在模式识别阶段比以前的方法获得更高的精度。