In this paper we propose to build large-scale datasets for learning local descriptors via mining recurring visual patterns. Such recurring patterns are discovered via partitioning correspondence graphs where each node represents a correspondence (i.e., similar keypoint pair) and each edge represents a geometric affinity between a correspondence pair. We evaluate our approach on a publicly available dataset and demonstrate that our approach is more efficient than an existing recurring pattern mining method. We also show that a descriptor learned from the datasets with our approach achieves superior performance on an image retrieval task than existing descriptors.