【摘要】 传统因果分析方法主要是基于具有分布预设的概率模型,但动力系统通常是存在反馈的非线性系统,不适合采用概率方法进行分析.针对这一问题,该文提出了基于扰动的亚复杂动力系统因果分析方法,主要工作包括:(1)采用基因表达式编程的函数拟合方法对动力系统时间序列进行差分方程拟合,减免了关于数据分布模型的预设;(2)基于得到的拟合函数,通过对自变量的扰动来计算因变量的相应波动,提出了根据扰动和波动的数值关系来判断自变量和因变量之间因果关系的判断准则,并基于该准则提出了因果关系挖掘算法和挖掘结果可信度验证方法;(3)在合成数据和真实数据上进行了翔实实验,结果表明该文所提出的算法能挖掘出合理因果关系,在不同数据规模情况下能得到一致挖掘结果.与两种基于概率统计的因果分析方法进行了对比实验,结果表明当系统要素多于两个时,该文的算法仍然能够得到多个要素间正确的因果关系,而两种基于概率统计的方法则无法挖掘出正确的因果关系.更多还原

【Abstract】 Traditional causality analyzing methods are based on probability models with predefined distributions.However,as dynamic systems are usually non-liner systems with feedback loops,the probability methods are not suitable for analyzing dynamic systems.In order to deal with this problem,this study proposes a new method to analyze the causal relationship between elements in sub-complex dynamic system.Main contributions include:(1)this study uses Gene Expression Programming to regress dynamic systems’ differential equations,and thus,avoid predefining the probability distribution function of the data;(2)based on the differential function,calculates the fluctuate value of the response variable by perturbing the independent variables.This study judges the causal relationships between the independent variables and the response variable by analyzing the numerical relationship between the response variables’ fluctuate valueand the independent variables’ perturbation value.Based on the judging criterion,this study proposes causality mining algorithm and causality trustiness evaluation algorithm;(3)conducts experiments on synthesized and real datasets.These results show that the algorithm can find reliable causal relationships which accord with social and natural principles.The algorithm can get the same results with different data scales.This study conducts comparative experiments with two causality analyzing methods which are based on probability analysis.These results show that when the system contains more than two elements,our method can still obtain the correct causal results,while the two comparative methods cannot find the correct causal results.更多还原