[4Xin1] Poster session 2

[4Xin1-17] Differences in Link Prediction Results across Multiple Datasets for Proposal of Metric to Evaluate the Machine Learning Suitability of Graph Datasets

〇Yuichi Yano¹, Kenichiro Akagi¹, Yayoi Natsume^2,5, Masataka Kuroda^2,4, Mari Itoh², Chihiro Higuchi², Kenya Ueno³, Mariko Nio⁶ (1.NEC Corporation, 2.AI Center for Health & Biomedical Research, National Institutes of Biomedical Innovation, Health and Nutrition, 3.Toray Industries, Inc., 4.Discovery Technology Laboratories, Mitsubishi Tanabe Pharma Corporation, 5.Institute of Advanced Medical Sciences, Tokushima University, 6.Chugai Pharmaceutical Co., Ltd.)

Keywords:Knowledge Graph, Knowledge Graph Embedding, Link Prediction, Side effect, Polypharmacy side effect

Many patients are administrated multiple drugs at the same time. Predicting drug synergy and adverse reactions because of drug combination therapies is needed for patients. Graph machine learning could solve this because biology, disease and drug are interconnected, their relationships can be represented as a network. On the other hand, in graph machine learning, generally, even small changes in the input dataset can lead to significant changes in the evaluation metrics. Thus, although many knowledge graphs embedding (KGE) algorithms have been proposed today, we cannot select the best one by comparing even the same evaluation metrics. In this study, we predicted side effects due to drug combinations using multiple different graph structure datasets from the same data source, then compared the results. With this result, we will propose new metrics to assess the suitability for machine learning of graph datasets. This metric could help to select the best KGE algorithms.