Cross-domain recommendation (CDR) alleviates the cold start problem by transferring the user-item rating patterns from a dataset in a dense rating auxiliary domain to one in a sparse rating target domain, and has been widely studied in recent years. The clustering methods based on single-domain recommendation adopted by most CDR algorithms fail to effectively utilize overlapping information and sufficiently adapt to CDR, resulting in inaccurate clustering results. In CDR, graph convolution network (GCN) methods can fully utilize the associations between nodes to improve recommendation accuracy. However, GCN-based CDR often employs static graph learning for node embedding, ignoring the fact that user preferences may change with different recommendation scenarios, which causes poor model performance across different recommendation tasks and ineffective mitigation of data sparsity. To this end, a multi-layer recurrent GCN CDR model based on a pseudo-overlap detection mechanism is proposed. Firstly, by fully leveraging overlapping data based on the community clustering algorithm Louvain, a pseudo-overlap detection mechanism is designed to mine user trust relationships and similar user communities, thereby enhancing the adaptability and accuracy of clustering algorithms in CDR. Secondly, a multi-layer recurrent GCN consisting of an embedding learning module and a graph learning module is proposed to learn dynamic domain-shared features, domain-specific features, and dynamic graph structures. By conducting iterative enhancement of the two modules, the latest user preferences are obtained to alleviate data sparsity. Finally, a multi-layer perceptron (MLP) is employed to model user-item interactions and obtain predicted ratings. Comparative results with 12 related models across four groups of data domains demonstrate the effectiveness of the proposed method, with average improvements of 5.47%, 3.44%, and 2.38% in MRR, NDCG, and HR metrics respectively.