非线性因果模型辨识方法

IDENTIFICATION METHOD FOR NONLINEAR CAUSAL MODELS

近来,基于观测变量的因果模型辨识受到了较多关注。一般使用线性无环因果模型对数据生成过程建模,而实际上,许多因果模型包含非线性关系,使用纯线性方法求解是无效的。将线性模型泛化为非线性模型,提出一种两步骤的辨识算法,首先使用特征选择算法获得d分离等价类,然后使用非线性成对独立性测试为图中的边标注因果方向。实验结果验证了该算法的有效性,并表明其优于其他算法。

The identification of causal models based on observed variables has received much attention in the past. Linear acyclic causal models are usually used to model the data-generating process, but practically many causal relationships are more or less nonlinear, this raises the doubts to the usefulness of purely linear methods. In this paper, we generalise the basic linear model to nonlinear model, and propose a two-step identification method, which first uses feature-selection algorithm to obtain the d-separation equivalence class, and then uses nonlinear pairwise independence tests to mark the causal directions for edges in the image. Experimental results verify the validity of this algorithm and show that it outperforms other methods.