Fig. 1

Problem formulation of multi-task causal graph learning in biomedicine. There are two scenarios of causal graph learning: a single-task (patient) setting and b multi-task setting. Each patient has a unique causal graph and his or her own observational data matrix. There are three dimensions to the data: the number of variables in the causal graphs, the number of patients/tasks, and the number of observational data points for each patient. Ideally, to deliver personalized medicine, it is important to construct a personalized causal graph for each patient. This study treats learning the causal graph for each patient as a single task and concentrates on multi-task settings. Although conventional approaches treat each task independently, we propose a novel causal inference approach to extract knowledge shared across all given tasks and to use meta-learning to enable rapid adaptation to new tasks