MAYPL: Structural Representation Learning on Hyper-Relational Knowledge Graphs Enables Inductive Reasoning over New Entities and Relations

Forum topic · ✨步子哥 · 2025-11-27

Summary

MAYPL (Structure Is All You Need) is a knowledge graph representation learning framework for hyper-relational knowledge graphs (HKGs) that performs inductive inference over both new entities and new relations. Unlike transductive methods such as TransE, RotatE, StarE, and HyNT, which learn fixed embedding vectors tied to training entities and relations, MAYPL relies purely on graph topology. It combines a structure-driven initializer, which computes initial representations from co-occurring neighbors using position-aware projection matrices, with an attention-based neural message-passing module that computes fact-level messages and iteratively aggregates information across facts, entities, and relations while accounting for component-level roles (head, tail, primary relation, qualifier). Because no entity- or relation-specific parameters are stored, a trained model can be applied directly to an entirely different inference HKG with unseen entity and relation sets, without retraining or fine-tuning. Compared with prior inductive methods, G-MPNN handles only one-hop new entities, HCNet handles new entities but not new relations, and QBLP depends on textual features. Evaluated on 10 benchmark datasets (e.g., WD50K, WikiPeople-, WD20K100v1, WK-50) against roughly 40 baseline methods, MAYPL achieves state-of-the-art results on transductive and inductive link prediction, with ablation studies confirming the necessity of both structural initialization and attention-based message passing. Applications include dynamic knowledge bases like Wikidata and YAGO, news event graph construction, and biomedical knowledge discovery.

Overview

MAYPL (Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs) introduces the first framework that performs inductive link prediction on hyper-relational knowledge graphs (HKGs) while handling both new entities and new relations simultaneously.

Key points

Inductive reasoning over HKGs: After training on an HKG G = (V, R, H), MAYPL can be directly applied to an inference HKG G' = (V', R', H') where the entity and relation sets may be entirely different (V ⊄ V' or R ⊄ R'), with no retraining or fine-tuning.
Purely structural learning: MAYPL does not learn fixed embeddings for specific entities or relations. Instead, it learns a universal procedure for computing, propagating, and aggregating messages over graph structure — hence "Structure Is All You Need."

Framework components

1. Structure-driven initializer: Initial representations are computed by aggregating co-occurring entities and relations. Position-specific learnable projection matrices distinguish structural roles (head entity, tail entity, qualifier entity, primary relation, qualifier relation). 2. Attention-based neural message passing:

*Fact-level messages* are computed from all entities and relations participating in a hyper-relational fact (a primary triple plus qualifiers).
These messages are aggregated back onto entities and relations via attention weights, iteratively over multiple layers, incorporating component-level connectivity.

Because all parameters are generic functions rather than element-specific vectors, representations for unseen entities and relations are generated on the fly from their local structure.

Comparison with prior methods

| Method | Handles HKG | Inductive new entities | Inductive new relations | | :--- | :--- | :--- | :--- | | MAYPL | ✓ | ✓ | ✓ | | StarE, HyNT, HAHE | ✓ | ✗ | ✗ | | NaLP, RAM | ✗ (NRR) | ✗ | ✗ | | G-MPNN | ✗ (KHG) | ✓ (one-hop only) | ✗ | | HCNet | ✗ (KHG) | ✓ | ✗ | | QBLP | ✓ | ✓ (needs text) | ✗ |

Existing transductive methods (TransE, RotatE, StarE, HyNT) require all test entities and relations to appear in training. Some approaches also convert HKGs to knowledge hypergraphs (KHG) or n-ary relation representations (NRR), which loses structural information about the primary triple and qualifier dependencies. Prior inductive methods are limited: G-MPNN only handles new entities adjacent to known ones; HCNet and QBLP cannot handle new relations.

Experiments and results

Evaluated on 10 benchmark datasets spanning transductive HKG, inductive KG, and inductive HKG settings (e.g., WD50K, WikiPeople-, WD20K100v1, WK-50).
Compared against approximately 40 baseline methods, achieving state-of-the-art performance on most tasks, with particularly large gains on inductive link prediction (measured by MRR and Hits@N).
Ablation studies confirm that both the structure-driven initializer and the attention mechanism are essential: replacing the initializer with random learnable vectors, or removing attention, causes significant performance drops in inductive settings.

Significance

Real-world knowledge bases like Wikidata and YAGO continuously gain new entities and relation types. A model requiring retraining for each new element is impractical. MAYPL's structural approach enables immediate integration and reasoning over new knowledge — relevant to scenarios such as building knowledge graphs from breaking news events and biomedical knowledge discovery involving novel genes, proteins, and drug interactions.