Link Prediction in Chemical Compound Network Under Observation Bias

Abstract

Data-based prediction of interactions between compounds is expected to have various applications including drug discovery. In recent years, there have been many attempts to predict compound networks by machine learning. However, there is a concern that various decisions made by chemists in the past regarding the selection of experimental targets may cause bias in the data used for learning, which in turn may lead to a decrease in prediction accuracy. In this study, in order to learn while correcting for this observation bias, we aim to improve prediction accuracy through representation learning using the HSIC, which is used as a measure of independence between random variables, as a regularization term. Experiments using semi-artificial data, in which observation bias is introduced to mimic experimental bias in real data, show that the proposed approach mitigates the bias and improves the prediction accuracy.