Knowledge graph-based entity alignment with unified representation for auditing

Abstract

Auditing is facilitated by audit knowledge graphs, while the biggest challenge in constructing an audit knowledge graph is entity alignment. Entity alignment involves linking entity pairs with the same real-world identity and aims to integrate heterogeneous knowledge across different knowledge graphs. However, most existing works do not effectively combine both attribute and relation representations into a unified framework for entity alignment, which is essential to link entities within an audit knowledge graph accurately. In this study, we propose a knowledge graph-based entity alignment approach with multi-attribute and weighted-relation fusion (KG-Marfia) for intelligent auditing. Our proposed KG-Marfia first extracts entity representations by addressing the imbalance of attributes and relations, and then designs a stacked graph convolutional network as an encoder to fuse attribute and relation information, learning unified representations for entities. In particular, we adopt an SVM-based classifier for the alignment task in intelligent auditing. Experiments conducted on two public datasets, as well as three audit datasets, demonstrate that our KG-Marfia outperforms state-of-the-art entity alignment methods.