SOHL: Semi-Ordered Hyperedge Representation Learning for Knowledge Hypergraphs

Traditional hypergraph representation learning for knowledge hypergraphs typically treats hyperedges as unordered collections of entities, failing to capture the critical sequential dependencies and directional semantics embedded within them. Conversely, existing ordered methods rely on rigid global entity orderings, rendering them brittle to noise and positional randomness. To strike a balance without the heavy overhead of strict positional encoding, we introduce SOHL (Semi-Ordered Hyperedge Learning), a novel representation learning framework centered on semi-ordered hyperedges. Rather than enforcing a rigid global order, SOHL partitions exchangeable entities into latent logical roles and retains only their valid sequential relations via a Counterfactual-guided Logical Role Mining Mechanism. Theoretical analysis further demonstrates both the strong expressive power and computational efficiency of our semi-ordered formulation. Extensive experiments across 16 benchmark datasets show that SOHL sets new state-of-the-art performance—outperforming top baselines by up to 3.2% in MRR—while demonstrating exceptional robustness against positional perturbations.