A Relative Feature Selection Algorithm for Graph Classification Advances in Databases and Information Systems

Graph classification is one of the most important research fields in data mining nowadays and many algorithms have been proposed to address this issue. In practice, labeling large or even medium-size graphs is a hard task and we need experts to do so. The biggest challenge in graph classification is extracting a set of proper features from graphs. Since graphs are represented by a complex data structure, this issue has been dealt with for a long time. Previous methods focused on extracting features from a certain class in a dataset. In this paper we propose a new feature selection method that extracts features from each graph rather than extracting them from a certain class in the dataset. We extract only frequent subgraphs as features. These subgraphs are chosen according to their number of occurrences in a graph. Moreover, we proposed a new formula which calculates the minimum number of occurrences required for a subgraph to be considered as frequent.We experimented on five real datasets and reached 7-17% higher accuracy than previously proposed methods.