Dissimilarity-based graph embedding transforms a graph into a vector representation, using a dissimilarity measure to capture relevant information between graphs. Dissimilarity-based graph embedding has been widely applied to supervised classification in social network analysis, bioinformatics, and biomedicine. Moreover, graph patterns have been used for dissimilarity-based graph embedding, since they represent regularities within the data. Mining graph patterns follows two main approaches: exact and approximate. In this paper, we compare the performance for supervised classification of dissimilarity-based graph embeddings built from exact versus approximate graph patterns. Our evaluation was conducted on four benchmark graph collections using different supervised classifiers, with accuracy and F1-score as performance metrics. The study reported in this paper allowed us to determine which kind of graph patterns, exact or approximate, allows building the dissimilarity-based graph embedding that achieves the highest supervised classification performance.