Targeted End-to-End Knowledge Graph Decomposition

Abstract

Knowledge graphs are networks with annotated nodes and edges, representing different relations between the network nodes. Learning from such graphs is becoming increasingly important as numerous real-life systems can be represented as knowledge graphs, where properties of selected types of nodes or edges are learned. This paper presents a fully autonomous approach to targeted knowledge graph decomposition, advancing the state-of-the-art HINMINE network decomposition methodology. In this methodology, weighted edges between the nodes of a selected node type are constructed via different typed triplets, each connecting two nodes of the same type through an intermediary node of a different type. The final product of such a decomposition is a weighted homogeneous network of the selected node type. HINMINE is advanced by reformulating the supervised network decomposition problem as a combinatorial optimization problem, and by solving it by a differential evolution approach. The proposed approach is tested on node classification tasks on two real-life knowledge graphs. The experimental results demonstrate that the proposed end-to-end learning approach is much faster and as accurate as the exhaustive search approach.