基于区域结构探测与边缘辨识发现时间序列因果关系转换

打开文本图片集

中图分类号：TP391 文献标志码：A

Discovering Causal Transitions in Time Series Based on Region Structure Detection and Edge Identification

XIE Jiea， WANG Kaijuna，b， FANG Yinga，b， LUO Tianjiana， b （a.Collegeof Computer and Cyber Security，b.Digit Fujian Internet-Things Laboratory ofEnvironmental Monitoring， FujianNormal University，Fuzhou 35O117，Fujian，China）

Abstract：Toimprove theaccuracyof existing methods for mining causal relationships over time，forabinarytime series containing one causal region，amethod was proposed tominethecausal transition pointsof timeseries byidentifying differentregional structuresandtheedgesof structure.Themethod designedthe posiblepositions ofthe causal region in timeseries as left，right，and center structure，used the existing causalitydiscovery methods to detect therough causal region，anddistinguisheditasacertainregional structureaccording tothediferentcharacteristicsof theregionstructures.According to the characteristics of diffrentregional structures，coresponding edge identificationmeasures were designed，and gradualy increasing detection windows andcausality intensity indexes were setto identify regional structural edges as causal transition points，and improvethe identification accuracyofcausal transition points.Experimentson two simulated datasetsand tworeal datasetsverified theaccuracyof the proposed method inrecognizing causal transition points.The resultsshow thatthe average acuracyofcausal transition pointsobtained bythe proposed method using Grangercausality scores onseparable simulated datasets is higher than those of the comparison methods，the average auracyof causal transition points obtained by convergent cross mapping causality scores on weakly coupled simulated datasets is higher than those of thecomparison methodatcoupling degresof O.O1and O.5O，and the accuracyof causal transitionpoints obtained byusing Grangercausalityscoresontworeal datasetsis higher than thatof thecomparison method.

Keywords：timeseries；causalityelation；causalrelation transition；convergentcross maping；Grangercausalitydetection

挖掘时间序列的因果关系，理解变量之间的因果关系对于预测、制定决策和解决问题至关重要[①]分析时间序列因果关系的常用方法包括Granger因果检验[2-3]和收敛交叉映射因果检测（CCM）[4]Granger因果检验和CCM方法在探索时间序列的因果关系方面的侧重点不同：对于可分离系统和紧密度很高的耦合系统，Granger因果检验方法有效；对于紧密度较低的弱耦合系统，CCM方法有效。（剩余10383字）