This paper introduces LaGR, a novel approach for integrating global information in knowledge graph reasoning that compresses the graph into a fixed, compact set of latent summaries and applies exact self-attention within this latent space, resulting in scale-invariant attention and stable performance across diverse data settings.
Knowledge graphs have become a fundamental representation for structured knowledge, yet their incompleteness remains a major obstacle to reliable reasoning. Missing links may interrupt relational paths and limit the evidence available for downstream inference. Knowledge graph completion addresses this problem by predicting missing facts from observed relational patterns. Existing methods have made notable progress, but many of them rely on a single geometric space or fuse multiple geometric representations in a static manner. This limits their ability to capture heterogeneous graph structures, especially when hierarchical dependencies, local relational patterns, and cyclic motifs coexist in the same knowledge graph. In this paper, we propose multi-curvature progressive fusion (MCPF), a geometry-aware graph neural reasoning framework for knowledge graph completion. MCPF learns relation-aware representations in Euclidean, hyperbolic, and hyperspherical spaces, which provide complementary inductive biases for conventional relational interactions, hierarchical organization, and cyclic structures. Instead of directly aggregating these representations, MCPF progressively refines them through cross-space interactions and uses a query-conditioned gating mechanism to determine their relative contributions. This design enables the model to emphasize reliable geometric signals while suppressing less informative ones for each query. The fused representation is then used within a graph neural reasoning architecture for message passing and link prediction. Experiments on four benchmark datasets demonstrate that MCPF achieves competitive results against representative baselines and consistently outperforms its ablated variants.
FITTER consistently outperforms inductive baselines without retraining, indicating that vocabulary-agnostic structural learning is a viable foundation for inference over the heterogeneous knowledge graphs of the Semantic Web.
Jia-Xin Pan, M. Nayyeri, Osama Mohammed et al.· 0 citations
