SLogic: Subgraph-Informed Logical Rule Learning for Knowledge Graph Completion

TL;DR

SLogic tackles KG completion by replacing static rule confidences with a query-dependent score $\phi(h, r, \mathbf{r_b})$ that integrates a head-centered subgraph via a Relational Graph Convolutional Network and a GRU-based rule encoder. A static rule base is mined using DFS up to length $L$ and reinforced with Wilson score lower bounds to form robust priors; during training, positive and hard negative rule pairs are created to learn contextual relevance. At inference, candidate rules are re-ranked using the context score, and entities are scored through a grounded, weighted aggregation that applies a tanh-squashed path count to mitigate hub effects. Experiments on WN18RR, FB15k-237, and YAGO3-10 show competitive or superior performance relative to both embedding-based and rule-based baselines, with notable gains on graphs with fewer relation types and clear explainability via explicit, query-specific rule grounding.

Abstract

Logical rule-based methods offer an interpretable approach to knowledge graph completion by capturing compositional relationships in the form of human-readable inference rules. However, current approaches typically treat logical rules as universal, assigning each rule a fixed confidence score that ignores query-specific context. This is a significant limitation, as a rule's importance can vary depending on the query. To address this, we introduce SLogic (Subgraph-Informed Logical Rule learning), a novel framework that assigns query-dependent scores to logical rules. The core of SLogic is a scoring function that utilizes the subgraph centered on a query's head entity, allowing the significance of each rule to be assessed dynamically. Extensive experiments on benchmark datasets show that by leveraging local subgraph context, SLogic consistently outperforms state-of-the-art baselines, including both embedding-based and rule-based methods.

Figures (4)

• Figure 1: Steps to generate instances enriched by context aware rules and to train SLogic model
• Figure 2: Algorithm to answer a query $\mathbf{q}(h,r,?)$
• Figure 3: Sensitivity analysis of MRR across three datasets. The optimal hyperparameters and the feasibility of using deep subgraph contexts vary significantly with KG structure. Higher number of hops could not be run for FB15k-237 and YAGO3-10 on our machine. )
• Figure 4: Ablation study of inference components and effect of inference hyperparameters. The ✗ symbol denotes that the component is disabled.