Abstract:
Graph-structured data are widely used in many fields of study. For example, in cheminformatics, the data of biochemical compounds can be encoded as a compound graph where atoms are represented by vertices and chemical bonds between pairs of atoms are represented by edges. By this setting, we can predict the molecular characteristics of an unseen biochemical compound by considering the similarity between two compound graphs. This problem is called a graph classification. This study, proposes an algorithm for compound graph classification by finding primitive substructures of the compounds and encode them in a new graph form. Then, their similarities are computed and fed to a support vector machine. The proposed algorithm was evaluated on four real-world data sets with four similarity measures, yielding the average accuracy of 84.7\%. Furthermore, the performance of the algorithm was also better than the classification method that solely uses graph kernels. In conclusion, the extraction of primitive structures can be a great tool for extracting compound structure features in biochemical compound classification.
