The graph classification is an important challenging task with many applications such as predicting the activity of chemical compounds. The presence or absence of subgraph patterns is often used as features for constructing a classifier. In this paper, we propose a graph classification method considering co-occurrences of subgraphs. This corresponds to an extension of existing linear methods to models including interaction terms of subgraph features. Subgraph indicators are 0-1 variables that

more »

... ve strong correlations due to the subgraph isomorphism between occurring subgraphs, and existing sparse linear models can be improved by taking these nonlinear interactions into account. We also present several variations to reduce candidate features increased by considering co-occurrence with some experimental verifications.