http://swrc.ontoware.org/ontology#TechnicalReport
Learning DAGs Graph Model from Sparse Inverse Covariance
en
Department of Information Science and Technology, The University of Tokyo
Department of Information Science and Technology, The University of Tokyo
Ziyu Guo
Shin Matsushima
Learning DAG-like causal graphs derived from structural equation models (SEMs) using data without intervention is one of the most important tasks in causal inference. However, in the real world, given observed data, the SEMs might be non-identifiable which makes the problem become hard. Recently, some studies have focused on inference on both identifiable and non-identifiable SEMs. These studies perform exhaustive search on all possible causal orders or set constraints on DAG. In this paper, we focus on learning SEMs without identifiability and propose an approach that estimates DAG computed by Cholesky decomposition of permutated precision matrix with BIC criterion. We perform graphical lasso to estimate sparse inverse covariance, which also controls the sparsity of the graph within l1 norm. Besides, we perform exhaustive search on all possible causal structures and use BIC criterion to construct the estimated DAG. Our experiments validate the performance of our approach on the SEMs without identifiability.
Learning DAG-like causal graphs derived from structural equation models (SEMs) using data without intervention is one of the most important tasks in causal inference. However, in the real world, given observed data, the SEMs might be non-identifiable which makes the problem become hard. Recently, some studies have focused on inference on both identifiable and non-identifiable SEMs. These studies perform exhaustive search on all possible causal orders or set constraints on DAG. In this paper, we focus on learning SEMs without identifiability and propose an approach that estimates DAG computed by Cholesky decomposition of permutated precision matrix with BIC criterion. We perform graphical lasso to estimate sparse inverse covariance, which also controls the sparsity of the graph within l1 norm. Besides, we perform exhaustive search on all possible causal structures and use BIC criterion to construct the estimated DAG. Our experiments validate the performance of our approach on the SEMs without identifiability.
AA12055912
研究報告バイオ情報学（BIO）
2022-BIO-70
25
1-6
2022-06-20
2188-8590