Causal Direction Inference Based On Observation Data

Inferring causal direction from observation data is a critical issue in data modeling for its further applications such as prediction and decision making. This step has been ignored in many circumstances because the causal direction can be occasionally obtained by naive ways such as expert knowledge, time order and experiment under control. But in many research field such as economics, ecology and physiology, when the above ways are no longer applicable, the only solution is to infer them from the observation data.The target of this essay is to discover the correct causal direction between two variables from observation when they are causally correlated. The main research content include:1) Causal direction inference method in nonlinear EIV system: In the EIV systems, the noises are symmetric which means that the only information can be used in causal direction inference is the nonlinear transfer function. In the existing methods, the useful information has always been buried in the observation noise when the signal-noise ratios are low. In this paper, a new algorithm is proposed to inferring the causal direction by estimate the explicit form of the transfer function and the distribution of the real input and output signals under the disturbing of the noises. The novel algorithm gives better results for large noises. 2) Causal direction inference in linear systems. In this part, the conditions for successful causal direction inference are discussed. The conditions are given by the inequalities which describe the relationship between the system transfer parameter and signal-noise ratio. Based on this discussion, a method based on the concept of unit entropy is proposed in the part. The method has been extended to solve causal sorting problem. 3) The causal direction inference algorithm based on non-parametric model is discussed. The users do not have to figure out the explicit model structure when implementing the non-parametric method. This is di?erent from the former two methods. And in order to get faster results when calculating the equivalent vector of transfer functions,recursive algorithm for Gaussian process model inference is developed. 4) The proposed methods in this paper are used to solve the propagation direction problems in network alert and atmosphere quality data set. The algorithms developed in this paper give same causal directions with the expert knowledge in the above fields.