多元知识图谱链接预测研究综述

Summary

Many real-world facts involve more than two entities. One example is a person who studied a subject at a university and earned a degree there. Splitting such facts into triples through auxiliary nodes, such as Freebase's CVT entities, loses structural information, adds parameters and makes reasoning harder. N-ary knowledge graphs (NKGs) store these facts directly, but like ordinary KGs they are incomplete. This survey presents itself as the first systematic review of NKG link prediction, the task of recovering missing elements of n-ary facts. It sorts nearly 50 published models along three axes: modeling strategy, fact representation and application setting. For each family, the authors explain the design idea, its strengths and its weaknesses. They also compile benchmark datasets, compare reported results of leading models, review applications in biomedicine and product recommendation, and outline future research directions. The paper gives a structured overview of a fast-growing area that had no dedicated survey before.

Key points

• Motivation statistics from Freebase: more than a third of entities take part in n-ary facts, and over 61% of fact types are n-ary.
• The surveyed methods fall into three modeling families:
  • Space-embedding: Euclidean (m-TransH, the first NKG link prediction model, and RAE), hyperbolic (e.g., HYPER2, PolygonE) and complex-space (ShrinkE, HYPERMONO).
  • Tensor decomposition: e.g., GETD, HypE, S2S.
  • Neural networks: FCN (NaLP, NeuInfer), CNN (HINGE, HyConvE, HJE), Transformer (GRAN, HyNT, NYLON) and GNN (StarE, HAHE, HyperFormer). Neural models form the most active line of work.
• Three fact formats are compared: hyperedge, unordered role–value pairs, and hyper-relational (a main triple plus qualifiers). The hyper-relational format is the most widely adopted because it stays compatible with ordinary triples.
• Besides the general static setting, the survey covers specialized scenarios: temporal (NE-NET, HypeTKG), few-shot (HANCL, MetaRH) and inductive (QBLP, MetaNIR).
• The authors propose three criteria for choosing NKG modeling over plain triples: participation of several entities, tight semantic coupling between them, and dependence on context such as time or place.