Traditional new drug development is difficult due to the burden of time and cost, and the risk of failure. Drug repurposing has the advantage that the success rate is relatively higher because it starts with a low risk related to drug safety or pharmacokinetics. With the advent of heterogeneous biomedical large-scale data, drug repurposing through in silico has become possible, but there is a difficulty in properly utilizing the heterogeneous characteristics of the data. In this paper, I propose a knowledge graph-based drug repurposing model, DRRE, that utilizes heterogeneous characteristics of biomedical data. The DRRE model learns topological features through GCN and reconstructs eight interactions. In particular, the heterogeneous characteristics of interactions are taken into account by introducing embeddings for each relation type. DRRE performed better than other heterogeneous graph-based models in predicting drug-disease interactions, and the fact that the DRRE performance was higher than the excluded model proved the effectiveness for relation type-specific embedding. In predicting drug candidates for Alzheimer’s disease conducted as a case study, three out of the top ten candidates predicted by DRRE have demonstrated efficacy in the literature.