River of No Return: Graph Percolation Embeddings for Efficient Knowledge Graph Reasoning

Kai Wang,Siqiang Luo,Dan Lin

2023-05-17

Abstract:We study Graph Neural Networks (GNNs)-based embedding techniques for knowledge graph (KG) reasoning. For the first time, we link the path redundancy issue in the state-of-the-art KG reasoning models based on path encoding and message passing to the transformation error in model training, which brings us new theoretical insights into KG reasoning, as well as high efficacy in practice. On the theoretical side, we analyze the entropy of transformation error in KG paths and point out query-specific redundant paths causing entropy increases. These findings guide us to maintain the shortest paths and remove redundant paths for minimized-entropy message passing. To achieve this goal, on the practical side, we propose an efficient Graph Percolation Process motivated by the percolation model in Fluid Mechanics, and design a lightweight GNN-based KG reasoning framework called Graph Percolation Embeddings (GraPE). GraPE outperforms previous state-of-the-art methods in both transductive and inductive reasoning tasks while requiring fewer training parameters and less inference time.

Artificial Intelligence

What problem does this paper attempt to address?

The paper is primarily dedicated to addressing the issue of path redundancy in Knowledge Graph (KG) reasoning and proposes a new method to improve the efficiency and accuracy of KG reasoning.

Specifically, the paper addresses the following issues:

1. **Theoretical Analysis of Path Redundancy**: The paper is the first to link the path redundancy problem with transformation errors in model training. It theoretically analyzes the entropy increase caused by path redundancy in path encoding schemes and provides a clear definition of redundant paths.

2. **Proposing Graph Percolation Embeddings (GraPE)**: To overcome the path redundancy problem in traditional graph propagation processes, the authors designed a KG reasoning framework called GraPE based on Graph Neural Networks (GNN). This framework maintains the shortest paths and removes redundant paths through a graph percolation process inspired by percolation models in fluid mechanics, thereby achieving efficient message passing.

3. **Experimental Validation**: Extensive experiments were conducted on multiple datasets, including transductive and inductive reasoning tasks, demonstrating that GraPE significantly improves reasoning performance while reducing computational complexity, especially excelling in inductive reasoning tasks.

In summary, through theoretical analysis and innovative method design, the paper effectively addresses the issue of path redundancy in KG reasoning, enhancing the efficiency and accuracy of KG reasoning models.