Graph neural network (GNN)–based collaborative filtering has shown strong performance in recommender systems by exploiting high-order user–item interactions. However, existing methods often suffer from noisy message propagation, uniform neighbor aggregation, and over-smoothing, particularly under sparsity. To address these limitations, we propose ChamGRec, which jointly models user–item interactions, item–item sequential relations, and meta-semantic relations within a unified GNN architecture. The proposed approach introduces a lightweight attention-based aggregation mechanism that computes adaptive neighbor-importance weights via a learnable linear attention function, enabling the model to emphasize informative neighbors while suppressing weak or noisy signals, without requiring explicit neighbor-selection procedures that would incur linear computational overhead. Furthermore, a cluster-guided user embedding adjustment strategy is employed to preserve user-specific preference distributions and mitigate over-smoothing in deeper propagation layers. Additionally, we design a secondary graph integration scheme that incorporates sequential and semantic knowledge into core collaborative representations at the embedding level. Extensive experiments on three real-world benchmark datasets: Amazon-Books, Yelp, and MovieLens-1 M, demonstrate that ChamGRec consistently outperforms state-of-the-art GNN-based recommender models in terms of Recall and NDCG, particularly in sparse settings, while effectively reducing noise diffusion and maintaining embedding discriminability. The source code and other data will be available at https://github.com/jhansilakshmivigrahala/ChamGRec-.