Relation prediction in knowledge graphs: A self-organizing neural network approach

Abstract

Knowledge graphs (KGs) in specialized domains frequently suffer from incomplete information. While current relation prediction methods for KG completion typically rely on neural network-based representation learning, we present KG2ART—a novel self-organizing neural network that employs a fundamentally different approach. KG2ART performs parallel inference over the graph structure through bidirectional interactions between bottom-up activations and top-down pattern matching to conduct relation prediction without representation learning. Our comprehensive evaluation across five diverse KGs (Nations, UMLS, Kinship, CoDEx-M, and a jet engine technical KG) demonstrates that KG2ART consistently outperforms state-of-the-art baselines (TuckER, ComplEX, RESCAL, ConvE, CompGCN) in prediction accuracy. The model achieves particularly strong results on standard benchmarks, with Hits@1 scores exceeding 90% for Nations and 60% for CoDEx-M. Remarkably, KG2ART attains these superior accuracy results while also being among the fastest models for both training and prediction across all datasets.