Graphs are a powerful representation formalism for structural data. They are, however, very expensive from the computational point of view. In pattern recognition it is often necessary to match an unknown sample against a database of candidate patterns. In this process, however, the size of the database is introduced as an additional factor into the overall complexity of the matching process. To reduce the influence of that factor, an approach based on machine learning techniques is proposed in this paper. Graphs are represented using feature vectors. Based on these vectors a decision tree is built to index the database. The decision tree allows at runtime to eliminate a number of graphs from the database as possible matching candidates.