大数据下高维交互效应的序贯仿真变量识别及应用

Sequential simulation variable identification and application of high-dimensional interaction effects for big data

随着大数据的到来,因子间的高维交互效应识别成为热门研究问题.为了解决该问题,对高效的Morris理论框架(EMM)进行扩展,提出了一种新的两阶段交互效应筛选方法(IEMM).IEMM充分利用计算机序贯特性,不仅可以高效识别出具有显著的主效应和交互效应和/或非线性效应的因子,而且可以进一步识别出因子间显著的交互效应(TIE).蒙特卡罗仿真结果表明,IEMM方法可以正确识别出EMM无法识别的TIE,并且与传统的识别交互效应的MM方法(IMM)相比,IEMM识别TIE的计算效率更高.实际物流案例表明IEMM识别因子间交互效应可行且稳健,并且能够在不牺牲统计效力的情况下节省计算资源.

Since the advent of big data,the identification of high-dimensional interaction effects between factors has become an active research area.To tackle the problem,this paper proposes a new two-stage interaction-based EMM method(IEMM)by extending the efficient Morris method-based framework(EMM).IEMM makes full use of the sequential characteristics of computers.Not only can IEMM efficiently identify factors with significant main effects and interaction and/or non-linear effects,but it can also further discern the significant two-way interaction effects(TIE).Monte Carlo simulation results show that IEMM can correctly identify TIEs that EMM cannot recognize.Compared to the traditional interaction-based MM(IMM),IEM-M has higher computational efficiency in TIE recognition.The real-world logistics case shows that IEMM is feasible and robust to identify TIEs and can achieve computational savings without sacrificing statistical effectiveness.