Research On Causality In Complex System Based On Bayesian Network

Causality is an ubiquitous and important relationship in complex systems. Causal discovery is an important means for people to see nature through phenomenon, and it can help people understand and master the internal causality of the specific domain.The causality in the complex system is often uncertain, indirect and implicit.Due to the existence of the causal deviation in the environment caused the system to appear causal inconsistency, it is difficult to find causality in complex systems.So that the causal transmission mechanism is more complex.From a system point of view, energy is a kind of decisive essential elements in a complex system. The surface characteristic of the system is the external manifestation of the energy, the event or the situation of the system is the result of the energy function. So as to make the energy become the link between the characteristics and the system trend,then energy analysis is the key to the discovery of causal relationship in the complex system.In this paper, the Causal Bayesian Network (CBN) as causal knowledge representation model,and the data mechanism made certain assumptions, thus causal discovery problem is transformed to learn model structure from a dataset. On this basis, the research work on the discovery of causality and its propagation in complex system is carried out.The main research contents include the following aspects:(1) Causal structure discovery of complex systems based on causal power and intervention test.Because the causality in the complex system is uncertain and weak, the causality between the elements in the system is difficult to discovery. Intervention learning through manipulating the parameters of the target node, observing the impact, and determining the causal link in the system, is an effective method of active learning.In view of the existing intervention learning method in selecting the intervention node does not consider the causal relationship between nodes and uncertainty and strength. Firstly, this paper introduces the causal mutual information to measure the strength of the causal relationship between the nodes. Then causal mutual information and asymmetric entropy are combined as causal uncertainly evaluation criteria to select the manipulated node and to produce intervention data. Finally, the causal structure discovery algorithm based on the causality and the intervention tests is proposed. Experimental results show that the proposed method can obtain the exact causal structure with less number of times of the intervention, and the learning effect is better than the existing methods.(2) Study on the causal propagation process of complex system based on sensitivity analysis.The cause and effect in the complex system is deep. And the influence of the change of the essential factors will spread and accumulate in the system, even the butterfly effect will appear.Due to the propagation of causal relation and the deep of causal effect in complex system, the effect of the parameters on the causal propagation is analyzed by sensitivity function. Taking the sensitivity as the basis of the parent node selection, the selection mechanism of the causal path in the complex system is studied.Firstly, the relative concepts of Bayesian network sensitivity analysis are given. And then describes the causal transmission process based on sensitivity analysis. Finally, the causal chain search algorithm based on sensitivity analysis is proposed, and the performance and complexity of the algorithm are analyzed.(3) Study on the motive mechanism of causal propagation of complex system based on energy calculation. Energy is an essential factor in the complex system, and the event is the result of the energy function. So the energy is the motive power of the system’s causal relationship.The stock market system is the dynamic and complex system of non-linearity and high noise.The fluctuation of stock market trend is essentially a process of energy change, so the thought based energy can realize the effective forecast of stock market trend.In this paper, firstly, the energy characteristics of the complex system are analyzed and then the energy elements of the stock market trend are extracted, and the energy calculation model of different characteristics is given, and the energy distribution is analyzed. Secondly, based on the feature fusion of Bayesian network, the energy propagation process is analyzed and the structure of stock market trend structure is constructed.Finally, the constraint relationship between the probability function of the energy condition is introduced into the support vector machine, and an energy forecast algorithm based on energy calculation for stock market situation is proposed. The experiment uses the three-year data of SSE (Shanghai Stock Exchange) for comparison and analysis.The experimental results show that the ideas based on energy can effectively solve the problem of inconsistency of index, and make the accuracy of situation forecast to be effectively improved.