How to find all overlapping communities in a complex network? That is, how to find all relevant groups of nodes in a linked dataset? No entirely satisfying solution to that important problem exists, having a criterion to decide which group is relevant and finding quickly these groups in large networks are bottlenecks. We found that in many networks the number of these groups is limited and that there exist, for each group, at least one node that can characterize it by itself: a node belonging only to that group and important within it. We call such a node a community representative. We develop an algorithm to find these overlapping communities. The community detection is done through measuring the proximities of all nodes to the representatives and then finding irregularities in the decrease of these values reflecting the presence of relevant groups. We show that our approach handles very large real-world networks and have comparable or even better performances compared to state of the art methods.