An integrated graph neural network for supervised non-obvious relationship detection in knowledge graphs

Abstract

Non-obvious relationship detection (NORD) in a knowledge graph is the problem of finding hidden relationships between the entities by exploiting their attributes and connections to each other. Existing solutions either only focus on entity attributes or on certain aspects of the graph structural information but ultimately do not provide sufficient modeling power for NORD. In this paper, we propose KGMatcher– an integrated graph neural network-based system for NORD. KGMatcher characterizes each entity by extracting features from its attributes, local neighborhood, and global position information essential for NORD. It supports arbitrary attribute types by providing a flexible interface to dedicated attribute embedding layers. The neighborhood features are extracted by adopting aggregation-based graph layers, and the position information is obtained from sampling-based position aware graph layers. KGMatcher is trained end-to-end in the form of a Siamese network for producing a symmetric scoring function with the goal of maximizing the effectiveness of NORD. Our experimental evaluation with a real-world data set demonstrates KGMatcher’s 6% to 35% improvement in AUC and 3% to 15% improvement in F1 over the state-of-the-art.