作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (1): 93-98. doi: 10.19678/j.issn.1000-3428.0060176

• 人工智能与模式识别 • 上一篇    下一篇

基于级联加性噪声模型的因果结构学习算法

乔杰1, 蔡瑞初1, 郝志峰2   

  1. 1. 广东工业大学 计算机学院, 广州 510006;
    2. 佛山科学技术学院 数学与大数据学院, 广东 佛山 528000
  • 收稿日期:2020-12-03 修回日期:2021-01-23 发布日期:2021-01-12
  • 作者简介:乔杰(1993-),男,博士研究生,主研方向为数据挖掘、机器学习;蔡瑞初(通信作者)、郝志峰,教授、博士生导师。
  • 基金资助:
    国家自然科学基金(61876043,61976052)。

Causal Structure Learning Algorithm Based on Cascade Additive Noise Model

QIAO Jie1, CAI Ruichu1, HAO Zhifeng2   

  1. 1. School of Computer, Guangdong University of Technology, Guangzhou 510006, China;
    2. School of Mathematics and Big Data, Foshan University, Foshan, Guangdong 528000, China
  • Received:2020-12-03 Revised:2021-01-23 Published:2021-01-12

摘要: 现有级联非线性加性噪声模型可解决隐藏中间变量的因果方向推断问题,然而对于包含隐变量和级联传递因果关系的因果网络学习存在全局结构搜索、等价类无法识别等问题。设计一种面向非时序观测数据的两阶段因果结构学习算法,第一阶段根据观测数据变量间的条件独立性,构建基本的因果网络骨架,第二阶段基于级联非线性加性噪声模型,通过比较骨架中每个相邻因果对在不同因果方向假设下的边缘似然度进行因果方向推断。实验结果表明,该算法在虚拟因果结构数据集的不同隐变量数量、平均入度、结构维度、样本数量下均表现突出,且在真实因果结构数据集中的F1值相比主流因果结构学习算法平均提升了51%,具有更高的准确率和更强的鲁棒性。

关键词: 因果结构学习, 加性噪声模型, 级联加性噪声模型, 因果发现, 函数式因果模型

Abstract: The existing cascade nonlinear Additive Noise Model(ANM) can infer the causal direction of hidden intermediate variables, but fail to deal with global structure search and equivalence class recognition in the case of causal network learning that includes hidden variables and cascade causality transferring.This paper presents a two-stage causal structure learning algorithm for non-chronological observation data.In the first stage, a basic causal network skeleton is constructed based on the conditional independence between the observation data variables. In the second stage, by using a cascaded nonlinear ANM, the causal direction of the edge likelihood under the assumptions of different causal directions is inferred by comparing each adjacent causality in the skeleton.The experimental results show that the algorithm has outstanding performance on the virtual causal structure dataset for a varying number of hidden variables, average in-degree, structural dimension, and number of samples.Furthermore, the F1 value of this algorithm on the real causal structure dataset improved by 51% on average compared with mainstream causal structure learning algorithms, displaying a higher accuracy and robustness.

Key words: causal structure learning, Additive Noise Model(ANM), Cascade Additive Noise Model(CANM), causal discovery, functional causal model

中图分类号: