Likelihood-based causal structure learning method

Abstract

The invention provides a likelihood-based causal structure learning method. The method mainly comprises the steps of performing directed edge addition, deletion and overturning processing on initialized structure graphs D; calculating score values SGk of the casual structure graphs by utilizing observation data; selecting the causal structure graph corresponding to the maximum score value MaxSGk;comparing an increased value of the maximum score value of the causal structure graph with a threshold epsilon; and through multi-time iteration, obtaining a final causal structure graph. By effectively fusing likelihood and structure equation search, the problem of incapability of identifying a causal structure due to existence of a Markov equivalence class in a conventional method is solved; a structure equation model is combined with the likelihood, so that the method can be applied to high-dimensional causal structure search; and by utilizing an xgboost classification algorithm and a kernel density estimation method, the application range of the causal structure learning method is further expanded, so that the method can be applied to linear or nonlinear data.

Description

technical field [0001] The invention relates to the technical field of data mining, in particular to a likelihood-based causal structure learning method. Background technique [0002] With the progress of society and the development of science and technology, the things people need to know become more and more complex. The causal relationship within the system exists objectively. The causal structure learning is to mine the causal structural relationship contained in the data, which can help people understand the complex The nature and laws of things. Causal structure learning has penetrated into various disciplines such as biology, medicine, economics, automatic control, and information processing, and involves various aspects such as daily life, industrial production, and military defense. [0003] The distribution of many variables in real life is often non-Gaussian. For example, the value of the magnetoencephalogram source does not necessarily conform to the standard Ga...

AI Technical Summary

Problems solved by technology

[0006] The current classical causal structure learning methods cannot effectively deal with continuous data with linear and arbitrary distributions with flow characteristics. The main limitations of these methods include:

[0007] 1. Most of the structural learning algorithms for linear arbitrary distribution are methods based on dependency analysis. In order to judge whether two features are related, this method needs to perform independence tests on a large number of subsets, resulting in the need for more independence tests. Thus It takes a lot of time, the calculation complexity is relatively large, and due to the nonlinear nature of the data, the calculation results of many nonlinear data are inaccurate;

[0008] 2. The structure learning algorithm for linear arbitrary distribution generally assumes that all data can be obtained in advance, and cannot process data with flow characteristics, that is, the features flow in one by one, so it cannot effectively deal with the causal structure learning problem in the dynamic and unknown feature space

[0009] 3. In the traditional structure learning, the structure cannot be recognized due to the existence of Markov equivalence class

Images