Learning Rule-Induced Subgraph Representations for Inductive Relation Prediction

TL;DR

This work addresses inductive relation prediction by distinguishing target links from irrelevant subgraph information. It introduces REST, a single-source edge-wise GNN that initializes only the target edge and uses RNN-based edge-wise message passing to encode rule-induced subgraphs, yielding representations focused on the target relation. Theoretical results show REST can learn rule-induced subgraph representations, and empirical results demonstrate state-of-the-art performance and significant subgraph preprocessing speedups (up to 11.66×). REST also offers interpretability through analysis of top-rule cycles, and its simplicity and efficiency make it a promising component for scalable reasoning in evolving knowledge graphs.

Abstract

Inductive relation prediction (IRP) -- where entities can be different during training and inference -- has shown great power for completing evolving knowledge graphs. Existing works mainly focus on using graph neural networks (GNNs) to learn the representation of the subgraph induced from the target link, which can be seen as an implicit rule-mining process to measure the plausibility of the target link. However, these methods cannot differentiate the target link and other links during message passing, hence the final subgraph representation will contain irrelevant rule information to the target link, which reduces the reasoning performance and severely hinders the applications for real-world scenarios. To tackle this problem, we propose a novel \textit{single-source edge-wise} GNN model to learn the \textbf{R}ule-induc\textbf{E}d \textbf{S}ubgraph represen\textbf{T}ations (\textbf{REST}), which encodes relevant rules and eliminates irrelevant rules within the subgraph. Specifically, we propose a \textit{single-source} initialization approach to initialize edge features only for the target link, which guarantees the relevance of mined rules and target link. Then we propose several RNN-based functions for \textit{edge-wise} message passing to model the sequential property of mined rules. REST is a simple and effective approach with theoretical support to learn the \textit{rule-induced subgraph representation}. Moreover, REST does not need node labeling, which significantly accelerates the subgraph preprocessing time by up to \textbf{11.66$\times$}. Experiments on inductive relation prediction benchmarks demonstrate the effectiveness of our REST. Our code is available at https://github.com/smart-lty/REST.

Figures (3)

• Figure 1: Relevant and irrelevant rules.
• Figure 2: An overview of REST. REST organizes the single-source initialization method and the edge-wise message passing method in a unified framework to learn relevant rules representations within the subgraph for the target link. Different relevant rules are shown in different colors in part 3.
• Figure 3: Comparison between conventional message passing framework developed by GraILteru2020inductive and our REST. First, REST initializes node and edge features with single-source initialization. Then, REST employs Update function to update edge features. Finally, REST directly uses the embedding of the target link as the final subgraph representation, rather than the pooling of all node embeddings.

Theorems & Definitions (3)

• Definition 1: Rule-induced subgraph representation.
• Theorem 1
• Theorem 2