<p>Knowledge graphs are multi-relational data that contain massive entities and relations. Recently, graph neural networks have been reported outstanding performance in modeling knowledge graphs. However, most existing methods based on graph neural networks are limited by expensive labeled information and high-time-space complexity for large-scale Multi-relational Knowledge Graphs (MKGs). In this paper, we propose the Multi-relational Knowledge Graph Contrastive Learning (MKGCL) method, an end-to-end framework with contrastive learning to solve label problems, which considers the local structure of subgraphs and alleviates algorithm complexity. Firstly, MKGCL extracts relational subgraphs according to the relation types of MKGs. The node representations are learned with a graph neural network encoder, and the representations of different relational subgraphs are obtained by pooling related node representations. Secondly, contrastive learning is used to take fully advantage of multi-relational data and heterogeneous structures for MKGs. MKGCL contrasts node-level and subgraph-level embeddings to capture more structural information in MKGs. Moreover, the number of relational subgraphs for model training has a vital impact on time and space complexity for MKGs. By learning small-size samples, the MKGCL method achieves well results while reducing algorithm complexity. Finally, extensive experiments on four benchmark datasets demonstrate that MKGCL yields better link prediction performance than existing methods.</p>