Research On The Estimation Of Causal Effect With Variational Auto-Encoder

With the development of big data technology,many algorithms have been explored to mine this data.However,most of them are limited to mining the association relationship in the data without deep understanding of the causal relationship,so often lead to some wrong conclusions.Such mistakes are often unpredictable and dangerous,especially in the medical field or business decisions.Artificial intelligence algorithms based on association mining have such unavoidable flaws.Foresight scholars have already begun research on causal analysis,and they believe that this is the right path to strong artificial intelligence.At present,causal analysis mainly includes two parts:causal network discovery and causal effect estimation.The former are the theoretical basis of the latter,and the latter is the application and verification of the formerThe traditional causal network discovery method is prone to equivalence problems,that is to say,there exist an indiscriminate causal direction between some variables.In essence,the traditional causal network method based on independence tests can only identify the V-structure,and cannot distinguish the structure d-separated.Therefore,the two variables causal direction identification has attracted people's attention.They are mostly based on a certain hypothesis generation mechanism,which is very flexible and can handle more realistic scenarios.Currently,most of these causal function methods are based on the assumption of noise independence.The main idea are to use kernel estimations to measure the independence of noise,but this will lead to a sharp increase in the calculation of large sample data,and it is also easy to misjudge independence of noise.In order to break for the limitation of the additive noise assumption,I began to study the the causal direction identification of heteroscedastic data.At present,there are few studies on the algorithm of this scenario.I found that the fourth-order moment is a good tool for measuring the characteristics of heteroscedastic noise.Based on the study of the fourth-order moment of noise,I have established a framework of causal direction identification of heteroscedastic data.The innovation of this algorithm is that the Gaussian process of variational heteroscedasticity can reduce the computational complexity,also the fourth-order moment theory breaks for the assumption of noise independence and further expands the application scope of the causal direction identification.At the same time,I also theoretically prove that our model can be degraded to the scene of the noise independent assumption,which shows that the algorithm has general applicability.The results of the experiment show that our method is promisingThe estimation of causal effect is based on the causal network discovery.Causal effect estimation has a wide range of applications in real-life scenarios,such as drug therapy,social sciences,etc.In fact,many of our decision-making problems are a combination of a series of causal effect estimation problems.The traditional causal effect estimation methods are mostly aimed at a small amount of low-dimensional data,and perform poorly when processing high-dimensional massive data,which is an inevitable problem in the era of big data.With the rapid development of deep learning algorithms,its effectiveness in handling high-dimensional masses is remarkable,which allows us to see the possibility of combining the deep learning methods and traditional causal effect estimation methods.The work of causal effect estimation in this paper is based on the variational auto-encoder.Our method focuses on solving the problem of causal effect estimation with hidden variables in high dimensions.I found that based on the variational auto-encoder model,ones can separate the hidden variables that are really effective from the estimation of causal effect.In a sense,we can achieve the compression and extraction of high-dimensional data information.Based on the correct separation of hidden variables,I combine the traditional propensity scoring method to solve the problem of high-dimensional causal effect estimation.The innovation of this method is to combine the streaming variational encoder and the traditional propensity score method to successfully solve the problem of causal effect estimation with hidden variables in high dimensions.The experiments show that the method is effective in the estimation of the average causal effectMy work of this paper analyzes the difficulties of causal analysis from theory to application,solves the problems of heteroscedasticity and high-dimensional data encountered in practical applications,and completes the theoretical framework from causal discovery to causal effect estimation.