For federated learning systems, the Non-IID data distribution across multiple clients brings enormous challenges to the model’s generalization ability. Most existing solutions rely on simple statistical indicators or heuristic rules. Due to the lack of comprehensive capture and modeling capabilities of the potential relationships between client data distributions, they fail to effectively leverage the feature distribution relevance between clients to promote model training. In this paper, we propose a novel approach for personalized federated learning, FedCRA, equipped with an adaptive aggregation strategy. Specifically, we construct a client relevance-aware graph based on the underlying data distribution differences among clients and propose a collaborative graph-assisted personalized aggregation method. We conducted extensive experiments on multiple public datasets, validating the effectiveness and superior performance of our proposed method.