[4Xin1] Poster session 2

[4Xin1-18] Experiments on Dataset Split Methods for Link Prediction using Knowledge Graph Embeddings

〇Mariko Nio¹, Tomoyuki Namai¹, Ryouichi Chatani¹, Manabu Wada¹ (1.Chugai Pharmaceutical Co., Ltd.)

Keywords:Knowledge Graph, Knowledge Graph Embeddings, Link Prediction

Papers about the relationships between gene knockout/knockdown and disease are important to select drug target genes. Since biomedical entities are interconnected in a network, the graphical data can be used to support this. However, applying ML (machine learning) to KG (knowledge graph) needs graph-specific approach. For example, for general ML we split dataset into training/validation/test sets, while for graph ML we need to select a suitable for task and graph-specific split method. In this study, we predicted gene-disease links from a KG using two data split settings and three algorithms. The KG was built with biomedical papers which were searched from Elsevier Text Mining. The results varied depending on the experimental settings. No algorithm was consistently better, the metrics also varied depending on settings. This result suggests that the combination of method would change the result, researchers may need to take it into account in their own studies.