Cellular D2D networks consist of numerous D2D user equipments (UEs) carried by human beings with multiple social attributes, which accordingly connotes an overlapping community (OC) structure. Accurately detecting OC D2D UEs can effectively improve the efficiency of multi-hop D2D communication. Existing relay selection methods overlook the OC structure characteristic of cellular D2D networks, which results in limited relay efficiency. Therefore, we propose an OC deep exploring-based relay selection method. First, we build the social tie matrix between D2D UEs and then based on deep learning theory we extract the features of social tie matrix to further precisely detect OC D2D UEs. Moreover, by reasonably utilizing detected OC D2D UEs, we design the effective relay selection method. Simulation results demonstrate that the proposed method can largely enhance the delivery rate and power consumption performances of cellular D2D networks.