Bibliographische Detailangaben
| Titel: |
外生干预与可逆流驱动的弱监督因果表征学习模型. |
| Alternate Title: |
Weakly-supervised causal representation learning model driven by exogenous intervention and invertible flow. |
| Autoren: |
张起荣1 zqr_hk@163.com, 王彪1 |
| Quelle: |
Application Research of Computers / Jisuanji Yingyong Yanjiu. Nov2025, Vol. 42 Issue 11, p3340-3347. 8p. |
| Schlagwörter: |
*CAUSAL models, *COUNTERFACTUALS (Logic), *BIG data, *MATHEMATICS, *CAUSATION (Philosophy), *MACHINE learning, *SUPERVISED learning |
| Abstract (English): |
Causal representation learning is a pivotal technology for achieving interpretability and intervenability in complex systems. However, current research faces several challenges: linear assumptions struggle to capture nonlinear causal dependen-cies in high-dimensional data, scarce labeled data limits model generalization. Notably, current frameworks exhibit deficiencies in both controllable intervention mechanisms and counterfactual inference capacities. Leveraging the crucial role of exogenous variables in explaining causal relationships and supporting counterfactual reasoning, this paper proposed a weakly supervised causal representation learning model driven by exogenous intervention and invertible flow. Firstly, it introduced exogenous vari-ables to simulate intervention scenarios, with causal graphs visually presenting causal paths and dependencies to achieve con-trollable intervention and counterfactual reasoning. Secondly, it employed an invertible flow model to capture nonlinear causal dependencies, overcoming the limitations of linear assumptions. Building on this, it incorporated a dynamic weakly supervised alignment mechanism, utilizing limited labeled data to constrain the semantic identifiability of causal factors, thereby mitigating the issue of scarce annotations. Experimental results on the Causal3DIdent dataset demonstrate significant performance improve-ments; the model achieves 94.5% accuracy in causal factor identification (8.8% increase over baseline models) and reduces intervention mean squared error to 0.015(47.7% decrease). Additionally, on datasets such as Pendulum-v1, the model exhibits strong performance, particularly in scenarios with limited labeled data, enabling effective causal inference and showcasing promising generalization capabilities and application potential. [ABSTRACT FROM AUTHOR] |
| Abstract (Chinese): |
因果表征学习是实现复杂系统可解释可干预的关键技术。当前研究存在线性假设难以捕捉高维数据中的非线性因果依赖、标注数据稀缺限制模型泛化能力且缺乏可控干预与反事实推理能力的问题。借助外生变量在因果联系解释、反事实推理支持方面的重要作用, 构建了一种外生干预与可逆流驱动的弱监督因果表征学习模型。首先, 引入外生变量模拟干预场景, 通过因果图直观呈现因果路径与依赖关系, 实现可控干预与反事实推理。其次, 采用可逆流模型捕捉非线性因果依赖, 突破线性假设限制; 在此基础上引入动态弱监督对齐机制, 利用少量标注数据约束因果因子语义可识别性, 缓解标注数据稀缺问题。实验结果表明, 在 Causal3 DIdent 数据集上, 模型取得了显著的性能提升, 因果因子识别准确率达到 94.5% (较基线模型提升 8.8%),干预均方误差降低至 0.015 (下降 47.7%)。此外, 在 Pendulum- v 1 等数据集上, 该模型同样表现出较好的性能, 尤其在面向标签数据稀缺情况下仍能实现有效因果推断, 展现出良好的泛化能力与应用前景。 [ABSTRACT FROM AUTHOR] |
| Datenbank: |
Academic Search Index |
Nájsť tento článok vo Web of Science