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.
CITATION STYLE
Škrlj, B., Kralj, J., & Lavrač, N. (2018). Targeted End-to-End Knowledge Graph Decomposition. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11105 LNAI, pp. 157–171). Springer Verlag. https://doi.org/10.1007/978-3-319-99960-9_10
Mendeley helps you to discover research relevant for your work.