摘要
Learning a Directed Acyclic Graph (DAG) structure from observational data plays an essential role in causal inference and machine learning. A recent advance in the area is that the DAG learning problem was formulated as a continuous optimization problem and this provides a new research line for leveraging powerful neural networks for DAG learning. Although several DAG learning algorithms using neural networks (called neural DAG learning) have been proposed, they still face some serious challenges, leading to unsatisfactory performance. To tackle this issue, in this paper, we combine ideas from local DAG learning and neural DAG learning techniques and propose a new Skeleton-Neural DAG (SN-DAG) structure learning algorithm. The SN-DAG algorithm first learns the structure skeleton of each variable using a local learning algorithm to initialize the weighted adjacency matrix, then employ multilayer perceptrons to optimize the weighted adjacency matrix for learning a DAG structure, and finally it corrects wrong edges in the learnt DAG using the learnt skeleton and v-structure identifying techniques. By conducting experiments on both synthetic and real datasets, experimental results validate the SN-DAG algorithm compared to traditional DAG learning algorithms and neural DAG learning methods.
学科与关键词
- Adaptive Skeleton Construction for Accurate DAG LearningGuo, Xianjie; Yu, Kui*; Liu, Lin; Li, Peipei; Li, JiuyongIEEE Transactions on Knowledge and Data Engineering, 2023, 35(10): 10526-10539.被引 3
- A novel data enhancement approach to DAG learning with small data samplesHuang, Xiaoling*; Guo, Xianjie; Li, Yuling; Yu, KuiApplied Intelligence, 2023, 53(22): 27589-27607.
- Towards Efficient Local Causal Structure LearningYang, Shuai; Wang, Hao; Yu, Kui*; Cao, Fuyuan; Wu, XindongIEEE Transactions on Big Data, 2022, 8(6): 1592-1609.被引 8
