Knowledge Probing for Graph Representation Learning

TL;DR

GraphProbe presents a principled framework for interrogating graph representations by three knowledge probes—node-wise, path-wise, and structure-wise—to reveal what graph properties are encoded by a wide range of graph-learning methods. By formalizing problem definitions and evaluating nine methods across six benchmarks on node classification, link prediction, and graph classification, the study shows that centrality information and structural signals are variably captured, with GAT often excelling in node-centric tasks and GCN/WGCN delivering broad task versatility. The probes demonstrate alignment with traditional performance metrics in several settings and offer complementary insights for model selection and interpretability. This work advances practical understanding of how graph properties influence downstream performance and provides a benchmark framework for future probing of graph representations.

Abstract

Graph learning methods have been extensively applied in diverse application areas. However, what kind of inherent graph properties e.g. graph proximity, graph structural information has been encoded into graph representation learning for downstream tasks is still under-explored. In this paper, we propose a novel graph probing framework (GraphProbe) to investigate and interpret whether the family of graph learning methods has encoded different levels of knowledge in graph representation learning. Based on the intrinsic properties of graphs, we design three probes to systematically investigate the graph representation learning process from different perspectives, respectively the node-wise level, the path-wise level, and the structural level. We construct a thorough evaluation benchmark with nine representative graph learning methods from random walk based approaches, basic graph neural networks and self-supervised graph methods, and probe them on six benchmark datasets for node classification, link prediction and graph classification. The experimental evaluation verify that GraphProbe can estimate the capability of graph representation learning. Remaking results have been concluded: GCN and WeightedGCN methods are relatively versatile methods achieving better results with respect to different tasks.

Figures (3)

• Figure 1: The overall architecture of our proposed GraphProbe.
• Figure 2: Radar chart comparison of graph embedding models for different information embedding capabilitie
• Figure 3: Skyline Plot on different Downstream Tasks based on random initialization