Differential-Aware Synergy of Quantity and Topology for Hyper-Relational Knowledge Graphs with Numeric Entities
Representation learning on hyper-relational numeric knowledge graphs (HNKGs), which incorporate numeric entities and facts consisting of a primary triple augmented by attribute-value qualifiers, is crucial for advanced reasoning. However, most existing methods employ discretization or binning strategies, struggling to consider physical quantities in modeling numeric entities: neglecting numeric continuity (e.g., 100 is greater than 1) and numeric polysemy (e.g., 80 represent age or weight without unit), leading to feature entanglement between numeric semantics and topological structures. In this predicament, standard attention mechanisms—limited by the inherent nature of positive-only accumulation (i.e., non-negative weights)-fail to peel away quantitative noise, forcing models into a compromised balance. To address these challenges, we propose NumDAE—a novel differential-aware decouple-then-synergize embedding framework for HNKGs. We reconstruct numeric semantics based on the relational context to endow numeric entities with algebraic computability, effectively decoupling pure physical quantity from the entangled topological background. Simultaneously, we design a differential-aware embedding transformer to effectively aggregate the semantic and structural representations, utilizing a subtractive cancellation between signal and noise attention maps to avoid mutual interference arising from the misalignment, achieving a high-fidelity synergy. Extensive experiments on public HNKG benchmarks demonstrate that NumDAE outperforms competitive baselines on link, relation and numeric value prediction tasks.