[2Win5] Poster session 2

[2Win5-25] An Introduction of Numerical Information into Knowledge Graph Embeddings for Numerical Prediction Based on Link Prediction

〇Akito Watanabe¹, Kohei Makino¹, Yutaka Sasaki¹ (1.Toyota Technological Institute)

Keywords:Knowledge Graph Embedding, Numerical Prediction, Link Prediction, Materials Informatics

This paper proposes a knowledge graph embedding to enable the inference of numerical attributes attended to nodes through link prediction, which uses a knowledge graph with edges of magnitude relationships between nodes that have close numerical values. Link prediction is a framework for predicting relationships between entities in a knowledge graph, which is realized by knowledge graph embedding. Since existing knowledge graph embedding treats numerical information as node attributes, quantitative characteristics are not reflected in the embedding space. Our method enables numerical property prediction by link prediction reflecting the quantitative characteristics. Our Experiments using datasets for material property prediction in Matbench showed that numerical prediction based on link prediction has the potential to improve prediction performance.