Relation-aware context aggregation framework for sparse knowledge graph completion

The task of knowledge graph completion (KGC) is to infer missing links that exist between entities in knowledge graphs (KGs). However, real-world KGs are typically structurally sparse, resulting in insufficient neighborhood context and thus degraded KGC performance. Meanwhile, most existing methods adopt coarse-grained relation modeling and overlook meaningful semantic representations of relations. To tackle these issues, we propose relation-aware context aggregation (RACA) for sparse KGC. Specifically, we introduce a dual-mode relational learning mechanism that jointly learns from both within-relation and across-relation perspectives, capturing task-relevant dimensional semantics within each relation and implicit logical regularities across relations. A spatial gating unit is further designed to integrate these relational signals and generate expressive relation representations. Finally, we develop a dynamic relation attention network that leverages relation-aware features to dynamically guide contextual aggregation, enabling more effective context capture under sparsity. Extensive experiments on three benchmark datasets show that RACA achieves superior performance over state-of-the-art baselines, particularly in sparse scenarios.