A lot of constrained patterns (e.g., emerging patterns, subgroup discovery, classification rules) emphasize the contrasts between data classes and are at the core of many classification techniques. Nevertheless, the extremely large collection of generated patterns hampers the end-user interpretation and the deep understanding of the knowledge revealed by the whole collection of patterns. The key idea of this paper is to summarize the contrasts of a dataset in order to provide understandable characterizations of data classes. We first introduce a novel framework, called recursive pattern mining, for only discovering few as well as relevant patterns. We demonstrate that this approach encompasses usual pattern mining framework and we study its key properties. Then, we use recursive pattern mining for extracting k recursive emerging patterns. Taken together, these patterns form a REP k-summary which summarizes the contrasts of the dataset. Finally, we validate our approach on benchmarks and real-world applications on the biological domain, showing the efficiency and the usefulness of the approach.