Knowledge graphs have become a popular method for representing large, relational data. Similar to citation networks and social networks, relationships in chemical reaction data can also be uniquely captured using a knowledge graph. However, relatively few studies exist concerning the application of knowledge graph mining techniques for numerical representation of chemical reactions. In this study, we develop a pipeline for transforming large-scale relational databases of chemical reactions into heterogeneous graphs, in which reactions and their reactants and products are all characterized as nodes with connecting edges. We create nodes for reaction templates, each of which links to multiple reactions to enhance the connectivity of the graph, and then employ graph representation learning methods (Node2Vec and RotatE) to generate an embedding (or fingerprint) for each reaction node. To evaluate the efficacy of this method, we construct classifiers to label the mechanisms of reactions based on these fingerprints. Experimental results show that our graph learning approach outperforms the state-of-the-art reaction fingerprints, specifically when class labels are not available during the representation learning process. When the representations can be fine-tuned for the subsequent classification task, our approach achieves comparable accuracy to a recent Transformer-based algorithm, but with a significantly lower computational cost.