<p>Recently, heterogeneous graphs have attracted significant interest as effective instruments for modeling intricate systems characterized by varied node and edge characteristics. However, when node attributes are missing in heterogeneous graphs, it hinders effective representation learning, which may lead to inaccurate classification results and unreliable clustering structures. In heterogeneous graphs with missing attributes, existing attribute completion algorithms exhibit two primary limitations. First, existing heterogeneous graph methods mainly rely on metapath-based instances or coarse-grained relation types, while ignoring the number and composition of relation paths between specific node pairs, which limits their ability to capture fine-grained multi-hop heterogeneous relation semantics. Second, they indiscriminately aggregate neighborhood information and ignore noisy nodes that are topologically connected but semantically inconsistent, thereby introducing interference into attribute completion. In this paper, we propose an algorithm named AC-HRCNF (Attribute Completion based on Heterogeneous Relation Context aware and Noisy node Filtering) to address these issues. First, we design a Heterogeneous Relation Context-aware (HRC) mechanism that explicitly models relation-level instances by counting the types and numbers of one-hop and multi-hop paths, enabling more fine-grained heterogeneous relation semantics than metapath-based instance modeling. Second, we develop the Noisy node Filtering (NF) strategy to eliminate neighbors that exhibit semantic inconsistency or abnormal features, thus enhancing attribute completion and supporting subsequent graph learning. Experiments on four real-world heterogeneous networks demonstrate that AC-HRCNF consistently outperforms existing attribute completion methods, achieving average performance gains of up to 16.89% in node classification and 43.77% in node clustering when averaged across all baselines.</p>