Causal discovery of cascaded data based on adversarial variational method
Xia-peng Zhang, Yaping Wan, Lijun Yang, Jing‐Hua Yang · 2021 3rd International Academic Exchange Conference on Science and Technology Innovation (IAECST) · 2021
At present, causal reasoning mainly studies the direct causal effects between variables, and uses the observed data to identify the causal direction of variables. However, in real data, there may be intermediate variables between the cause and the result, which leads to indirect nonlinear causal effect between the initial cause and the final result, which makes the accuracy of traditional causal inference methods low. To solve this problem, based on the cascading nonlinear additive noise model, proposes a causal direction inference algorithm: adversarial variational cascade additive noise model (AVCANM). Based on adversarial variational method, which estimates the lower variational bound of the marginal log likelihood of the data through adversarial training, and then determine the causal direction. The feasibility of this method is proved in theory, and simulation experiments are carried out on simulated data and real data. The results of the experiment show that compared with the traditional causal inference algorithms, this method has better stability in identifying indirect and nonlinear causal relationship.