计算机与现代化 ›› 2025, Vol. 0 ›› Issue (11): 119-126.doi: 10.3969/j.issn.1006-2475.2025.11.015

• 算法设计与分析 • 上一篇    

去噪扩散概率模型驱动的因果推断算法

  


  1. (河海大学数学学院,江苏 南京 211100)
  • 出版日期:2025-11-20 发布日期:2025-11-24
  • 作者简介: 作者简介:张洪滔(2001—),男,山东滕州人,硕士研究生,研究方向:因果推断,深度生成式模型,E-mail: zht18763293829@163.com; 通信作者:朱永忠(1968—),男,江西瑞昌人,教授,博士生导师,研究方向:数据科学,随机系统分析,数理统计的理论及应用研究,E-mail: yzzhu@hhu.edu.cn; 张雨轩(2001—),女,河南南阳人,硕士研究生,研究方向:因果发现,表征学习,E-mail: 231362020009@hhu.edu.cn; 夏世源(1998—),男,安徽合肥人,博士研究生,研究方向:因果发现,因果森林,E-mail:syxia@hhu.edu.cn。
  • 基金资助:
     基金项目:国家自然科学基金资助项目(71831006)
      

Causal Inference Algorithm Driven by Denoising Diffusion Probabilistic Model


  1. (School of Mathematics, Hohai University, Nanjing 211100, China)
  • Online:2025-11-20 Published:2025-11-24

摘要: 摘要:针对传统因果推断算法未考虑干预措施对结构因果模型的影响,处理高维复杂因果关系数据时存在推断偏差较大、稳定性较低等问题,结合去噪扩散概率模型提出一种新型的因果推断算法。首先,分析不同干预类型对结构因果模型和扩散模型的影响,提升算法的可解释性。然后,从理论上推断因果马尔可夫假设下模型的似然下界,构造了包含因果父代的扩散采样过程。最后,结合基于扩散的因果模型的变分结构,可以使本文算法同时从原始数据和干预数据中进行采样,并对干预数据进行反事实估计,从而简化模型的训练复杂度,增强模型的鲁棒性。与其他算法进行对比,模拟结果表明,在不同类型的结构因果模型假设下,本文算法采样的最大平均差异降低了10%~41%,反事实估计的均方误差降低了9%~63%。实证结果表明,本文算法采样的最大平均差异降低了11%,反事实估计的均方误差降低了5%。实验表明,本文算法可以有效处理复杂的数据结构和噪声分布,以及显著提升采样和反事实估计的准确性和稳定性。



关键词: 关键词:结构因果模型, 神经网络模型, 去噪扩散概率模型, 随机过程, 反事实推断, 变分推断

Abstract: Abstract:A new causal inference algorithm is proposed to address the limitations of traditional methods, which fail to account for the impact of interventions on structural causal models and suffer from large inference bias and low stability when handling high-dimensional, complex causal data. This algorithm integrates a denoising diffusion probabilistic model. First, it analyzes the effects of different types of interventions on both the structural causal model and the diffusion model to improve algorithm interpretability. Next, the likelihood lower bound of the model under the causal Markov assumption is theoretically derived, and a diffusion sampling process that includes causal parents is constructed. Finally, by combining the variational structure of the diffusion-based causal model, the algorithm can sample from both original and intervention data and perform counterfactual estimation, thus simplifying model training complexity and enhancing robustness. In comparison with other algorithms, simulation results show that under different structural causal model assumptions, the proposed algorithm reduces the maximum mean discrepancy by 10%~41% and the mean squared error of counterfactual estimation by 9%~63%. Empirical results indicate a reduction of 11% in the maximum mean discrepancy and 5% in the mean squared error of counterfactual estimation. Experiments demonstrate that this algorithm effectively handles complex data structures and noise distributions, significantly improving the accuracy and stability of both sampling and counterfactual estimation.

Key words: Key words: structural causal model, neural network model, denoising diffusion probability model, stochastic processes, counterfactual inference, variational inference

中图分类号: