Presentation 1-3, 4:00PM~4:30PM JST.
Speaker
Sunmin Oh (Seoul National University)
Title
Learning Algorithms for Directed Acyclic Graphical Models.
Abstract
Directed acyclic graphical (DAG) models play a crucial role in modeling and understanding complex systems across diverse fields. This presentation explores various DAG learning algorithms, providing a comprehensive overview of their theoretical foundations. In addition, I highlight their strengths and limitations, focusing on the assumptions and sample sizes required to recover the true model. Following this, I introduce a new advancement, the optimal backward learning algorithm (OptBGSM), that achieves sample optimality using the best-subset-selection method. Through numerical experiments and real data analysis, I demonstrate that the OptBGSM performs significantly better than the other algorithms.