Federated recommendation enables collaborative model training without exposing raw user data, serving as a privacy-preserving learning paradigm. Despite recent advances, existing methods often fail to simultaneously achieve high recommendation accuracy and low communication overhead under data heterogeneity—a common characteristic in real-world scenarios. To address this, we propose ClusLazyFR, a federated recommendation algorithm that combines user clustering with adaptive lazy aggregation. By organizing users into clusters with similar behavioral patterns and sampling them in a more structured manner, ClusLazyFR promotes balanced participation, thereby improving convergence and accuracy. Building on this clustering, a dynamic lazy aggregation strategy selectively triggers communication rounds based on inter-round parameter differences, allowing the system to retain useful updates while skipping redundant ones. Experiments on three public datasets—MovieLens, Filmtrust, and Ciao—demonstrate that ClusLazyFR consistently outperforms baseline methods in both accuracy and communication efficiency, achieving faster convergence and over 30% reduction in communication cost.