Graph Stochastic Neural Process for Inductive Few-shot Knowledge Graph Completion

TL;DR

The paper tackles inductive few-shot knowledge graph completion (I-FKGC), a setting where both new relations and unseen entities appear at test time. It introduces Graph Stochastic Neural Process (GS-NP), a two-module framework consisting of a neural-process-based hypothesis extractor that models a distribution over shared hypotheses and a graph stochastic attention-based predictor that tests query triples against sampled hypotheses and yields an explanatory subgraph. The approach unifies neural processes with graph-based attention by deriving an ELBO-based objective that combines prediction, hypothesis KL regularization, and subgraph-structure KL terms, enabling end-to-end training. Experiments on three benchmarks show state-of-the-art performance and provide interpretable subgraphs that explain predictions, demonstrating both effectiveness and transparency for inductive reasoning in KGs.

Abstract

Knowledge graphs (KGs) store enormous facts as relationships between entities. Due to the long-tailed distribution of relations and the incompleteness of KGs, there is growing interest in few-shot knowledge graph completion (FKGC). Existing FKGC methods often assume the existence of all entities in KGs, which may not be practical since new relations and entities can emerge over time. Therefore, we focus on a more challenging task called inductive few-shot knowledge graph completion (I-FKGC), where both relations and entities during the test phase are unknown before. Inspired by the idea of inductive reasoning, we cast I-FKGC as an inductive reasoning problem. Specifically, we propose a novel Graph Stochastic Neural Process approach (GS-NP), which consists of two major modules. In the first module, to obtain a generalized hypothesis (e.g., shared subgraph), we present a neural process-based hypothesis extractor that models the joint distribution of hypothesis, from which we can sample a hypothesis for predictions. In the second module, based on the hypothesis, we propose a graph stochastic attention-based predictor to test if the triple in the query set aligns with the extracted hypothesis. Meanwhile, the predictor can generate an explanatory subgraph identified by the hypothesis. Finally, the training of these two modules is seamlessly combined into a unified objective function, of which the effectiveness is verified by theoretical analyses as well as empirical studies. Extensive experiments on three public datasets demonstrate that our method outperforms existing methods and derives new state-of-the-art performance.

Figures (10)

• Figure 1: (a) An illustration of inductive few-shot link prediction (I-FKGC) where new relations emerge with unseen entities simultaneously. (b) An example of the motivation behind our method where we try to test whether the query set aligns with the hypothesis extracted from the support set.
• Figure 2: The overall framework of our proposed model GS-NP, which consists of two major modules, namely, a neural process hypothesis extractor (including (b)) and a graph stochastic attention-based predictor (including (c), (d), and (e)). (a) We first extract enclosing subgraphs for all triples from the background knowledge graph. (b) We adopt the GNN encoder and NP encoder to model the joint distribution of the hypothesis. (c) After sampling a hypothesis (i.e., $z$) from the distribution, we inject the hypothesis into the graph structure by the hypothesis fusion module. (d) We apply the graph stochastic attention to identify a subgraph and feed it into the GNN encoder to get representation. (e) We compute the cosine similarity between the subgraph representation and the hypothesis to test whether the query aligns with the shared hypothesis extracted from the support set.
• Figure 3: The training process of GS-NP
• Figure 4: Ablation study results on NELL, ConceptNet, and WIKI datasets.
• Figure 5: Parameter analysis of $r$ on NELL, ConceptNet, and WIKI datasets.