Inductive Link Prediction on N-ary Relational Facts via Semantic Hypergraph Reasoning

TL;DR

This work tackles fully inductive link prediction on n-ary relational facts by introducing an n-ary semantic hypergraph and the NS-HART subgraph aggregator. NS-HART uses a two-stage Transformer-based message passing to capture intra-edge role interactions and multi-hop correlations across facts, enabling robust inductive generalization beyond fixed training entities. The authors provide score-function analysis and demonstrate superior performance on inductive benchmarks across TR-EF, TR-NEF, and PSR tasks, with comprehensive ablations validating the design choices. The approach offers a principled pathway for reasoning over rich, high-order relations in knowledge graphs and delivers publicly available code for reproducibility.

Abstract

N-ary relational facts represent semantic correlations among more than two entities. While recent studies have developed link prediction (LP) methods to infer missing relations for knowledge graphs (KGs) containing n-ary relational facts, they are generally limited to transductive settings. Fully inductive settings, where predictions are made on previously unseen entities, remain a significant challenge. As existing methods are mainly entity embedding-based, they struggle to capture entity-independent logical rules. To fill in this gap, we propose an n-ary subgraph reasoning framework for fully inductive link prediction (ILP) on n-ary relational facts. This framework reasons over local subgraphs and has a strong inductive inference ability to capture n-ary patterns. Specifically, we introduce a novel graph structure, the n-ary semantic hypergraph, to facilitate subgraph extraction. Moreover, we develop a subgraph aggregating network, NS-HART, to effectively mine complex semantic correlations within subgraphs. Theoretically, we provide a thorough analysis from the score function optimization perspective to shed light on NS-HART's effectiveness for n-ary ILP tasks. Empirically, we conduct extensive experiments on a series of inductive benchmarks, including transfer reasoning (with and without entity features) and pairwise subgraph reasoning. The results highlight the superiority of the n-ary subgraph reasoning framework and the exceptional inductive ability of NS-HART. The source code of this paper has been made publicly available at https://github.com/yin-gz/Nary-Inductive-SubGraph.

Figures (7)

• Figure 1: Three ways to represent the n-ary relational fact: "$Company_{A}$, $Company_{B}$, $Company_{C}$ cooperate in project $AIP$." Note that, our proposed n-ary semantic hypergraph is based on the key-value pair representation.
• Figure 2: Task Examples. In transfer reasoning tasks, models are trained on one graph and then predict missing links among candidate entities in another inference graph (with no entity overlapping). In pairwise subgraph reasoning tasks, models are given the merged neighborhood of a target entity and known entities of an incomplete fact to assess the likelihood that the target completes the incomplete fact.
• Figure 3: The overall structure of the proposed learning framework. To illustrate the message passing processes of NS-HART, we take updating embeddings of hyperedge $e_1$ and entity $v_1$ at iteration t as an example.
• Figure 4: Results of the TR-EF and TR-NEF tasks on the WD20K (100) V1 using NS-HART, along with ablation studies for "w/o message passing" and "w/o high-order relations".
• Figure 5: Performance (on WD20K (100) V1 of TR-EF tasks) of NS-HART with varied hops under different sampling scales.