Causal network inference method based on kurtosis

Abstract

The invention relates to a causal network inference method based on kurtosis. The method comprises the steps that exogenous variables are selected based on the kurtosis, a causal order is found layer by layer, and a cutting edge is tested by a cutting edge using the least square method, so that a complete causal network is output. The causal network inference method based on kurtosis is not sensitive to the degree of non-Gaussian of disturbance variables and has strong stability, can still be able to maintain a high recognition rate especially when the non-Gaussian is weak, and has low complexity, kurtosis calculation is needed for each variable itself, the method is a direct estimate way, and the causal network recognition accuracy is high.

Description

technical field [0001] The invention relates to the technical field of data mining, in particular to a kurtosis-based causal network inference method. Background technique [0002] Currently, discovering causal relationships from observational data has received extensive attention and has applications in many fields, such as neuroscience, economics, and epidemiology. In the absence of any prior knowledge, traditional causal discovery methods can only find Markov equivalence classes, and intervention experiments are needed to obtain a complete network, but in reality, many variables cannot be intervened. [0003] This problem can be solved well by the linear non-Gaussian acyclic model (LiNGAM) proposed by Shimizu et al., which is able to identify complete causal networks from only observation datasets and has been increasingly studied. [0004] There are three main ideas for the estimation of the linear non-Gaussian acyclic model (LiNGAM) model: [0005] The first is to tra...

AI Technical Summary

Problems solved by technology

[0009] However, the first two ideas are to convert the problem into a function optimization problem, and then use some optimization algorithms to solve it, but they are sensitive to the initial value and are prone to local convergence defects; for the third method, the current variable is compared with the rest The most independent of the residuals of all variables is an unreliable measure of exogenous variables;

[0010] Estimating the LiNGAM model using an approximate transformation method will lead to local optimal problems or unreliable independence problems, especially when the interfering variables are closer to the Gaussian distribution, it is more difficult to distinguish the causal relationship between variables

Images