CN112069450A - 基于凸集间交互投影的多对象结构方程模型计算技术 - Google Patents

基于凸集间交互投影的多对象结构方程模型计算技术 Download PDF

Info

Publication number
CN112069450A
CN112069450A CN202011067786.0A CN202011067786A CN112069450A CN 112069450 A CN112069450 A CN 112069450A CN 202011067786 A CN202011067786 A CN 202011067786A CN 112069450 A CN112069450 A CN 112069450A
Authority
CN
China
Prior art keywords
model
structural
variable
equation model
solving
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011067786.0A
Other languages
English (en)
Inventor
童乔慧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan Zhuxin Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan Zhuxin Technology Co ltd filed Critical Wuhan Zhuxin Technology Co ltd
Priority to CN202011067786.0A priority Critical patent/CN112069450A/zh
Publication of CN112069450A publication Critical patent/CN112069450A/zh
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/11Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
    • G06F17/12Simultaneous equations, e.g. systems of linear equations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/06Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063Operations research, analysis or management
    • G06Q10/0639Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393Score-carding, benchmarking or key performance indicator [KPI] analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/06Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063Operations research, analysis or management
    • G06Q10/0639Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06395Quality analysis or management
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/01Customer relationship services

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Mathematical Physics (AREA)
  • Educational Administration (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Tourism & Hospitality (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明“基于凸集间交互投影的多对象结构方程模型计算技术”,技术领域属于电子与信息类的应用软件技术。本发明求解多对象结构方程模型,分为三个步骤。(1)将多对象结构方程模型原始数据纵向叠放,利用基于配方约束的模型确定性算法统一求解,得到每个结构变量对应的观测变量的汇总系数。(2)将叠放的数据块按结构变量纵向剖分,分别采用评估模型的凸集间的交互投影算法求解,得到每个对象每个结构变量的评估分。(3)将上一步计算得到的评估分矩阵作为新的观测矩阵,按结构方程模型求解,得到每个对象的顾客满意度最终评估分。

Description

基于凸集间交互投影的多对象结构方程模型计算技术
技术领域
本发明属于电子与信息类的应用软件技术,具体是一种基于凸集间交互投影的多对象结构方程模型计算技术。
背景技术
(一)一般的结构方程模型与顾客满意度
FORNELL教授最先将结构方程模型(SEM)引入到顾客满意度测评[1-2]。SEM包括两个方程组 ,一个是结构变量之间的关系方程组,称为结构方程组;一个是结构变量与观测变量之间的关系方程组,称为观测方程组。图1是一个典型的中国顾客满意指数模型。
我们在Excel表上把观测数据列出来,观测次数按行排列,而变量按列排列。自变量在左侧,因变量在右侧。观测数据阵
Figure RE-70460DEST_PATH_IMAGE001
Figure RE-226635DEST_PATH_IMAGE002
等等都是已知的,星号代表行数。系数是未知的,因变量也是未知的。图2只列出了模型中最左侧的3个变量。
结构方程组包含 6个结构变量(隐含变量)
Figure RE-554848DEST_PATH_IMAGE003
Figure RE-163684DEST_PATH_IMAGE004
与 11个关系 (自变量作用的关系为
Figure RE-602756DEST_PATH_IMAGE005
,因变量之间的作用关系为
Figure RE-613437DEST_PATH_IMAGE006
),
Figure RE-112551DEST_PATH_IMAGE007
是残差变量,如式(1)所示。
Figure RE-208683DEST_PATH_IMAGE008
(1)
在一般情形下,结构变量不一定是5个,结构方程系数形式也可以不同于式(1) ,自变量的个数也可以多于1个。如果采用向量与矩阵记法进行一般描述,设因变量有
Figure RE-202178DEST_PATH_IMAGE009
个,将
Figure RE-332945DEST_PATH_IMAGE010
排成列向量,记为
Figure RE-2961DEST_PATH_IMAGE011
;自变量有
Figure RE-320810DEST_PATH_IMAGE012
个,将
Figure RE-367263DEST_PATH_IMAGE013
排成列向量,记为
Figure RE-352537DEST_PATH_IMAGE014
Figure RE-662295DEST_PATH_IMAGE011
的系数矩阵为
Figure RE-998599DEST_PATH_IMAGE009
阶方阵,记为
Figure RE-52006DEST_PATH_IMAGE015
Figure RE-672212DEST_PATH_IMAGE014
的系数矩阵为
Figure RE-152871DEST_PATH_IMAGE016
阶矩阵,记为
Figure RE-242050DEST_PATH_IMAGE017
;残差向量为
Figure RE-833569DEST_PATH_IMAGE018
,则结构方程组式(1)可以扩展为:
Figure RE-324593DEST_PATH_IMAGE019
(2)
SEM的结构变量是隐含的 ,不能直接观测,且其对应若干个观测变量。设一共有
Figure RE-976154DEST_PATH_IMAGE020
个观测变量,对每一个观测变量有
Figure RE-287050DEST_PATH_IMAGE021
个观测 ,在顾客满意指数分析中就是有
Figure RE-682259DEST_PATH_IMAGE021
个顾客的测评,这样我们手里的数据是一个
Figure RE-778522DEST_PATH_IMAGE022
矩阵。
结构变量与观测变量之间的作用关系也可以用方程表示,按作用的因果路径有两种表示方式。
设与自变量
Figure RE-600984DEST_PATH_IMAGE023
对应的
Figure RE-868018DEST_PATH_IMAGE024
个观测变量为
Figure RE-598076DEST_PATH_IMAGE025
,
Figure RE-266955DEST_PATH_IMAGE026
;与因变量
Figure RE-791477DEST_PATH_IMAGE027
对应的
Figure RE-545807DEST_PATH_IMAGE028
个观测变量为
Figure RE-79556DEST_PATH_IMAGE029
,
Figure RE-602941DEST_PATH_IMAGE030
。于是从观测变量到结构变量的观测方程组可以表达为:
Figure RE-547633DEST_PATH_IMAGE031
Figure RE-789258DEST_PATH_IMAGE032
(3)
Figure RE-329961DEST_PATH_IMAGE033
Figure RE-239011DEST_PATH_IMAGE034
(4)
反之,从结构变量到观测变量的观测方程可以表达为:
Figure RE-839757DEST_PATH_IMAGE035
Figure RE-834257DEST_PATH_IMAGE032
(5)
Figure RE-178651DEST_PATH_IMAGE036
,
Figure RE-942208DEST_PATH_IMAGE037
(6)
其中,
Figure RE-713855DEST_PATH_IMAGE038
Figure RE-946384DEST_PATH_IMAGE039
为载荷项。上面两式采用矩阵记法可以表为:
Figure RE-828889DEST_PATH_IMAGE040
(7)
Figure RE-181373DEST_PATH_IMAGE041
(8)
上面的式子和图形结合起来称为结构方程模型,有时也称为路径分析模型。本课题组对它们开展了深入研究,提出了基于配方约束的确定性算法,可以取代传统的协方差拟合算法(Linear Structure RELationship,LISREL)与偏最小二乘算法(Partial LeastSquare, PLS)。同时本课题组还提出了多层结构方程模型,见图3(一个多层结构方程模型的变量与路径图),并且解决了它们的算法问题。
(二)配方回归模型
为了使得本专利技术通俗易懂,我们通过数据结构图逐步讲解。
先浅说什么是回归。班级考试加总分是容易理解的。一个班有
Figure RE-655080DEST_PATH_IMAGE021
个(例如30个)学生,每个学生考了
Figure RE-93015DEST_PATH_IMAGE020
门功课(例如4门),就有了一个数据阵,有
Figure RE-310369DEST_PATH_IMAGE021
(30)行,
Figure RE-517360DEST_PATH_IMAGE020
(4)列,每列向量分别记为
Figure RE-161968DEST_PATH_IMAGE042
。现在要加总分,需要知道每门考试的满分,比如分别是150,120,100,150。这样直接把每个同学的考分相加,实际上每门功课的成绩占比或者说分量是不一样的。满分高的功课占比大一些,显得重要一些。如果我们事先把所有考分都化成了百分制,满分统统是100分,那么在汇总的时候,各门功课成绩就要乘以不同的系数,这个例子里就是要分别乘以1.5, 1.2, 1.0, 1.5。这个占比的分量或者说系数就是加权系数,记为
Figure RE-87198DEST_PATH_IMAGE043
。汇总以后得到的总分是一个向量
Figure RE-311506DEST_PATH_IMAGE044
,有
Figure RE-419008DEST_PATH_IMAGE021
(30)个数据,并且:
Figure RE-437780DEST_PATH_IMAGE045
这里
Figure RE-381465DEST_PATH_IMAGE044
是未知待求的,而加权系数
Figure RE-409464DEST_PATH_IMAGE043
是已知的。如果把向量
Figure RE-122205DEST_PATH_IMAGE042
排在一起成为一个矩阵记为
Figure RE-577457DEST_PATH_IMAGE046
,把系数
Figure RE-742859DEST_PATH_IMAGE043
排成向量记为
Figure RE-308970DEST_PATH_IMAGE047
,则上式可以简记为
Figure RE-610638DEST_PATH_IMAGE048
。考试加总分的数据结构图见图4(考试加总分的数据结构图)。
普通回归就是在上述汇总过程中,假定因变量
Figure RE-518683DEST_PATH_IMAGE044
是已知的,而加权系数
Figure RE-640222DEST_PATH_IMAGE043
是未知待求的。回归的数据结构图看起来和图4的一样,不同的是因变量与回归系数已知与未知颠倒了,见图5(一元线性回归数据结构图)。
回归系数怎么求,肯定有误差,要使得误差平方和最小,于是采用了最小二乘法则。如图6所示(线性回归的最小二乘法则)。
误差平方和最小从欧式距离的角度理解就是投影,于是普通回归的几何意义就是求
Figure RE-541182DEST_PATH_IMAGE021
维空间里的一个点(向量
Figure RE-697357DEST_PATH_IMAGE044
)到一个子空间的投影,这个子空间是由
Figure RE-25570DEST_PATH_IMAGE020
个列向量(
Figure RE-634406DEST_PATH_IMAGE049
)张成的。理解这个几何意义对于我们下面寻找评估模型的算法非常重要,如图7(线性回归最小二乘法则的投影几何意义)。
所谓配方回归,就是在上面的回归模型中,还要求回归系数之和为1,并且每一个回归系数都不小于0。于是配方回归模型可写作
Figure RE-73478DEST_PATH_IMAGE050
(9)
这里
Figure RE-84159DEST_PATH_IMAGE051
是因变量(观测向量),
Figure RE-832541DEST_PATH_IMAGE052
是自变量(设计阵),
Figure RE-928673DEST_PATH_IMAGE053
是回归系数,
Figure RE-640277DEST_PATH_IMAGE054
是误差向量。如果记
Figure RE-36623DEST_PATH_IMAGE055
,则约束条件可记为
Figure RE-175481DEST_PATH_IMAGE056
(10)
它是一般线性约束
Figure RE-290067DEST_PATH_IMAGE057
的特殊形式。
配方回归的实际含义是各因素百分比的分摊。比如
Figure RE-805362DEST_PATH_IMAGE044
是总的发行债券,
Figure RE-56215DEST_PATH_IMAGE058
是各发行公司的发行能力,那么
Figure RE-365974DEST_PATH_IMAGE059
是各发行公司承担的发行份额在总任务中所占百分比。在化学配方与药品配方中,
Figure RE-718589DEST_PATH_IMAGE044
是待配的药品总量,
Figure RE-771995DEST_PATH_IMAGE058
是各药品的重量,
Figure RE-611775DEST_PATH_IMAGE059
是各药品在混料中所占的百分比。回归模型的任务是从历史数据中推断出一个比较合适而折衷的配方:
Figure RE-623594DEST_PATH_IMAGE060
在配方回归模型中,不仅
Figure RE-181614DEST_PATH_IMAGE046
是已知的,而且
Figure RE-304291DEST_PATH_IMAGE044
也是已知的,这是与下面将要讨论的评估模型不一样的地方。配方模型是一个典型的二次规划问题,即在约束
Figure RE-998577DEST_PATH_IMAGE061
(即
Figure RE-446876DEST_PATH_IMAGE062
),求二次型
Figure RE-226613DEST_PATH_IMAGE063
(11)
的最小值。由于约束条件表示一个闭凸锥,二次型最小值总是存在的。当
Figure RE-621823DEST_PATH_IMAGE046
列满秩时,解是唯一的。调用优化问题或规划问题中程序可以解算这个模型。
统计学家从回归原理也对此模型提出算法,主要是使用Lagrange乘子原理与原地扫除算法。简单地说,就是将约束条件分解为两部分,一个是线性约束
Figure RE-216621DEST_PATH_IMAGE056
(12)
一个是符号约束
Figure RE-39083DEST_PATH_IMAGE064
(13)
先解线性约束回归模型
Figure RE-837275DEST_PATH_IMAGE065
(14)
若其解
Figure RE-36175DEST_PATH_IMAGE064
,则它就是
Figure RE-236212DEST_PATH_IMAGE066
的最终解。若
Figure RE-229576DEST_PATH_IMAGE066
有某分量为负,则可以证明
Figure RE-515064DEST_PATH_IMAGE066
的最终解必在约束边界上,即有某个或某些
Figure RE-252076DEST_PATH_IMAGE067
Figure RE-57352DEST_PATH_IMAGE067
,即在原模型中剔除了变量
Figure RE-487196DEST_PATH_IMAGE058
,如此继续回归。
需要说明的是,建立模型时,样本组数
Figure RE-259980DEST_PATH_IMAGE068
与自变量个数是随意的,回归系数
Figure RE-535104DEST_PATH_IMAGE069
应该为0,
Figure RE-709733DEST_PATH_IMAGE070
之和应为1。这是由本模型特点所决定的。
(三) 评估模型
我们先从实际工作提炼出模型。
质量评估工作是常见而又重要的。根据
Figure RE-44900DEST_PATH_IMAGE071
个母体的
Figure RE-304980DEST_PATH_IMAGE020
个指标的观测值,来给这
Figure RE-383794DEST_PATH_IMAGE071
个母体打个分,排个队,现在是司空见惯的事情。如产品质量评估,作品质量评估,演出质量评估,地区部门工作质量评估,教师授课质量评估,等等。问题在于怎样打分比较合理,这需要建立数学模型。
我们还是回顾图5的数据结构图。评估模型里不仅回归系数要满足配方回归条件,而且因变量是未知的。如此而已,看样子并不复杂,但是因变量与回归系数都是未知的,那如何求得唯一解,原来这个模型里母体(班级个数)不止一个。为了适合表现现在的数据结构,我们改进图5为图8(评估模型数据结构图)。每个班级的评估分放在左边,一共有
Figure RE-662198DEST_PATH_IMAGE071
个班级,就有
Figure RE-433844DEST_PATH_IMAGE071
个数据块。
Figure RE-384483DEST_PATH_IMAGE020
个指标是变量,分别以
Figure RE-798147DEST_PATH_IMAGE072
表示。一张评估表是某一母体的一次观测,可取得数据
Figure RE-150631DEST_PATH_IMAGE073
。对
Figure RE-624337DEST_PATH_IMAGE074
个母体各取得
Figure RE-593430DEST_PATH_IMAGE075
次观测,就得
Figure RE-14047DEST_PATH_IMAGE046
阵。一张评估表是
Figure RE-768508DEST_PATH_IMAGE046
阵的一行,一个母体的
Figure RE-881957DEST_PATH_IMAGE021
次观测是
Figure RE-807188DEST_PATH_IMAGE046
阵的一块。对每个变量的加权系数
Figure RE-562654DEST_PATH_IMAGE076
待定,但需
Figure RE-624151DEST_PATH_IMAGE077
(即
Figure RE-439661DEST_PATH_IMAGE078
);
Figure RE-852187DEST_PATH_IMAGE079
(即
Figure RE-411345DEST_PATH_IMAGE080
)。这是一种配方约束。对每个母体必须且只须给出一个分数,它也是事先未知而待定的,这就是所谓广义。因此评估模型是如下三个式子联合组成。
Figure RE-592927DEST_PATH_IMAGE081
(15)
Figure RE-782600DEST_PATH_IMAGE056
(16)
Figure RE-474568DEST_PATH_IMAGE064
(17)
(15)(16)(17)三式合起来是一种广义配方模型(GP模型),它是杨自强研究的因变量可变的广义最小二乘模型与方开泰等研究的配方模型的结合。所谓广义,就是因变量未知。这里
Figure RE-40679DEST_PATH_IMAGE082
Figure RE-873505DEST_PATH_IMAGE083
Figure RE-234079DEST_PATH_IMAGE084
Figure RE-152357DEST_PATH_IMAGE085
Figure RE-522158DEST_PATH_IMAGE086
,
Figure RE-943912DEST_PATH_IMAGE087
,即
Figure RE-740967DEST_PATH_IMAGE088
。对
Figure RE-349803DEST_PATH_IMAGE071
Figure RE-805186DEST_PATH_IMAGE089
数据块按列分别求平均,得到压缩的数据阵
Figure RE-815868DEST_PATH_IMAGE090
下面先考虑GP模型中仅满足(15)、(16)的解。令
Figure RE-314982DEST_PATH_IMAGE091
(18)
Figure RE-411114DEST_PATH_IMAGE092
(19)
则由
Figure RE-653877DEST_PATH_IMAGE093
Figure RE-519064DEST_PATH_IMAGE094
Figure RE-189080DEST_PATH_IMAGE095
(20)
又由
Figure RE-772508DEST_PATH_IMAGE096
,令
Figure RE-802650DEST_PATH_IMAGE097
,不难验证
Figure RE-787924DEST_PATH_IMAGE098
为投影阵。记p维矩阵
Figure RE-97682DEST_PATH_IMAGE099
,当
Figure RE-699565DEST_PATH_IMAGE100
可逆时,
Figure RE-752971DEST_PATH_IMAGE066
解为
Figure RE-123910DEST_PATH_IMAGE101
(21)
总结上述过程,有
定理1.若
Figure RE-604570DEST_PATH_IMAGE102
,则在约束
Figure RE-428169DEST_PATH_IMAGE056
Figure RE-285267DEST_PATH_IMAGE103
min
有唯一解(20)、(21)。如果
Figure RE-261444DEST_PATH_IMAGE104
各分量非负,则(20)、(21)也就是
Figure RE-913005DEST_PATH_IMAGE105
模型的解。
当按(21)解出的
Figure RE-958322DEST_PATH_IMAGE104
有分量为负时,要考虑模型
Figure RE-884690DEST_PATH_IMAGE105
的解的存在性、唯一性,有如下定理。
定理2 若
Figure RE-433483DEST_PATH_IMAGE106
,则
Figure RE-52683DEST_PATH_IMAGE107
模型有唯一解。若(21)中
Figure RE-319716DEST_PATH_IMAGE104
有分量为负,则
Figure RE-49775DEST_PATH_IMAGE107
模型的解
Figure RE-453074DEST_PATH_IMAGE108
一定有分量为0,并且
Figure RE-492443DEST_PATH_IMAGE108
的零分量是
Figure RE-246773DEST_PATH_IMAGE104
的分量之一。
证明 (9)可以改写为
Figure RE-983784DEST_PATH_IMAGE109
,集合
Figure RE-38328DEST_PATH_IMAGE110
是闭凸集,故存在唯一点
Figure RE-468172DEST_PATH_IMAGE111
满足(9)。由于
Figure RE-975377DEST_PATH_IMAGE112
列满秩,故由
Figure RE-516080DEST_PATH_IMAGE113
能唯一解出
Figure RE-690709DEST_PATH_IMAGE114
再考虑集合
Figure RE-25876DEST_PATH_IMAGE115
(22)
Figure RE-36688DEST_PATH_IMAGE116
(23)
显然是两个闭凸集,
Figure RE-115503DEST_PATH_IMAGE117
有界。由两个闭凸集间距离可达定理,存在
Figure RE-144638DEST_PATH_IMAGE118
Figure RE-650706DEST_PATH_IMAGE119
,这里
Figure RE-398082DEST_PATH_IMAGE120
表示距离,且已证
Figure RE-280588DEST_PATH_IMAGE121
唯一。于是问题转化为求一点到
Figure RE-164230DEST_PATH_IMAGE117
的最短欧氏距离,即方开泰等研究的PR模型。现
Figure RE-106778DEST_PATH_IMAGE122
,由该文中的定理1,本定理得证。
但是
Figure RE-325139DEST_PATH_IMAGE123
的求法并没有解决,这些留待下面统一给出计算方法。
再考虑对
Figure RE-745756DEST_PATH_IMAGE124
的约束,设
Figure RE-218326DEST_PATH_IMAGE125
(24)
其中
Figure RE-597354DEST_PATH_IMAGE126
均已知,也已去掉了多余约束。考虑模型
Figure RE-522585DEST_PATH_IMAGE127
(25)
这里
Figure RE-278051DEST_PATH_IMAGE128
,其余假定同前。
定理3. 若
Figure RE-605128DEST_PATH_IMAGE106
,则
Figure RE-420637DEST_PATH_IMAGE129
模型有唯一解,对于存在的
Figure RE-833164DEST_PATH_IMAGE130
Figure RE-877474DEST_PATH_IMAGE131
解的性质如定理2。对称地,对于存在的
Figure RE-59057DEST_PATH_IMAGE048
Figure RE-514309DEST_PATH_IMAGE044
解的性质也如定理2。
(四) 凸集间的交互投影算法
求一点
Figure RE-945290DEST_PATH_IMAGE132
到闭凸集
Figure RE-511401DEST_PATH_IMAGE015
之间的最短欧氏距离,若
Figure RE-78648DEST_PATH_IMAGE133
, 则可以称
Figure RE-704802DEST_PATH_IMAGE134
Figure RE-623079DEST_PATH_IMAGE135
Figure RE-727301DEST_PATH_IMAGE015
的投影。自然它有别于一点到子空间的投影。要求两个闭凸集
Figure RE-663902DEST_PATH_IMAGE136
之间的最短欧氏距离,可以使用交互投影法。
任取
Figure RE-460957DEST_PATH_IMAGE137
,求
Figure RE-600951DEST_PATH_IMAGE138
,使
Figure RE-774444DEST_PATH_IMAGE139
。对于
Figure RE-316283DEST_PATH_IMAGE134
,求
Figure RE-284239DEST_PATH_IMAGE140
,使
Figure RE-911530DEST_PATH_IMAGE141
。对于
Figure RE-357555DEST_PATH_IMAGE142
,求
Figure RE-770213DEST_PATH_IMAGE143
,使
Figure RE-909070DEST_PATH_IMAGE144
。对于
Figure RE-23656DEST_PATH_IMAGE145
,求
Figure RE-273372DEST_PATH_IMAGE146
,使
Figure RE-789804DEST_PATH_IMAGE147
。当
Figure RE-99563DEST_PATH_IMAGE148
时,停止迭代,完成计算。
上述迭代过程收敛的意思是:
Figure RE-701445DEST_PATH_IMAGE149
(26)
定理4.设
Figure RE-754852DEST_PATH_IMAGE150
两个闭凸集之一有界,则其交互投影的迭代过程收敛。
证明 因为对
Figure RE-375058DEST_PATH_IMAGE151
Figure RE-855718DEST_PATH_IMAGE152
所以数列
Figure RE-679318DEST_PATH_IMAGE153
单减有界,极限存在。
不妨设
Figure RE-536415DEST_PATH_IMAGE015
有界,则点列
Figure RE-761860DEST_PATH_IMAGE154
中存在子列
Figure RE-413421DEST_PATH_IMAGE155
Figure RE-989896DEST_PATH_IMAGE156
。对应
Figure RE-385105DEST_PATH_IMAGE100
中子列
Figure RE-215789DEST_PATH_IMAGE157
也是有界的,其中存在子列
Figure RE-303831DEST_PATH_IMAGE158
,当然有
Figure RE-102023DEST_PATH_IMAGE159
Figure RE-35344DEST_PATH_IMAGE160
为过点
Figure RE-704222DEST_PATH_IMAGE161
而与线段
Figure RE-25482DEST_PATH_IMAGE162
垂直的平面,因为
Figure RE-514232DEST_PATH_IMAGE015
为闭凸集,故
Figure RE-703774DEST_PATH_IMAGE015
全在
Figure RE-227159DEST_PATH_IMAGE163
一侧。当线段
Figure RE-922583DEST_PATH_IMAGE164
时,
Figure RE-164208DEST_PATH_IMAGE163
Figure RE-236070DEST_PATH_IMAGE165
Figure RE-394387DEST_PATH_IMAGE015
全在
Figure RE-526291DEST_PATH_IMAGE166
一侧。
又对
Figure RE-989634DEST_PATH_IMAGE151
Figure RE-334028DEST_PATH_IMAGE167
数列
Figure RE-363163DEST_PATH_IMAGE168
单减有界,极限存在,对上述
Figure RE-869231DEST_PATH_IMAGE169
,对应
Figure RE-616607DEST_PATH_IMAGE100
中子列
Figure RE-499113DEST_PATH_IMAGE170
有界,其中存在子列
Figure RE-133487DEST_PATH_IMAGE171
,当然还是
Figure RE-810456DEST_PATH_IMAGE172
。由于
Figure RE-779549DEST_PATH_IMAGE100
为闭凸集,
Figure RE-465746DEST_PATH_IMAGE173
为定点,
Figure RE-203895DEST_PATH_IMAGE174
,故有
Figure RE-317344DEST_PATH_IMAGE175
现在是平面
Figure RE-998041DEST_PATH_IMAGE177
全在
Figure RE-325117DEST_PATH_IMAGE178
一侧。于是
Figure RE-401613DEST_PATH_IMAGE150
分别在两平行平面
Figure RE-548560DEST_PATH_IMAGE178
Figure RE-107718DEST_PATH_IMAGE179
两侧,线段
Figure RE-289300DEST_PATH_IMAGE180
是公垂线,同时
Figure RE-275711DEST_PATH_IMAGE181
,故
Figure RE-175534DEST_PATH_IMAGE182
,又
Figure RE-272803DEST_PATH_IMAGE183
收敛,
Figure RE-308892DEST_PATH_IMAGE184
。证毕
根据定理4,求两个闭凸集之间的距离可以化为累次求一点到闭凸集间的距离。于是求解广义配方模型可以化为累次求解配方模型,求解凸约束广义配方模型可以化为累次求解一般凸约束模型,实际计算表明,收敛过程非常快。见图9(凸集间的交互投影算法示意图)。
对于给定的初值
Figure RE-748095DEST_PATH_IMAGE185
Figure RE-400793DEST_PATH_IMAGE186
是超平面
Figure RE-770594DEST_PATH_IMAGE187
上的一个点,需要求解
Figure RE-457928DEST_PATH_IMAGE188
,模型是:
Figure RE-254982DEST_PATH_IMAGE189
此时
Figure RE-644244DEST_PATH_IMAGE131
有一个凸约束即配方约束,图9中是向下的投影。我们根据配方回归方法可以求解。一旦我们解得
Figure RE-552157DEST_PATH_IMAGE131
的估计
Figure RE-93997DEST_PATH_IMAGE190
,
Figure RE-61953DEST_PATH_IMAGE191
就是凸集
Figure RE-689244DEST_PATH_IMAGE192
上的一个点,我们需要求得对应的
Figure RE-400848DEST_PATH_IMAGE193
的解,此时模型是:
Figure RE-62773DEST_PATH_IMAGE194
这是图9中向上的投影,我们按照普通回归求得解
Figure RE-483521DEST_PATH_IMAGE195
。如此反复迭代,参考文献证明了交互投影的收敛性。
只要理解了图6,知道回归就是求误差平方和最小;理解了图7,知道误差平方和最小就是投影,那么就可以理解图9,向
Figure RE-332529DEST_PATH_IMAGE196
的投影是无约束的,向
Figure RE-847824DEST_PATH_IMAGE197
的投影是有配方约束的,于是通过交互投影就可以求解模型。
主要参考文献目录
[ 1 ] Fornell C., Johnson M. D., Andrson E.W. , et al. The Americancustomer satisfaction index: nature, purpose, and findings[J ]. Journal ofMarketing, 1996, 60 (4) : 7 - 18.
[ 2 ] Claes F., A national customer satisfaction barometer: the Swedishexperience [ J ]. Journal of Marketing, 1992, 56 (1) : 6 - 21.
[ 3 ] 国家质检总局质量管理司, 清华大学中国企业研究中心. 中国顾客满意指数指南 [M ]. 北京 :中国标准出版社 , 2003: 21 - 58.
[ 4 ] Inon F. A., Llar Io R. Development of a PLS based method fordetermination of the quality of beers by use of N IR: spectral ranges andsample - introduction considerations[ J ]. Analytical and Bio analyticalChemistry, 2005, 382 (7) : 1549 - 1561.
[ 5 ] Tenenhausm Vinzive, Chatelin Y. M., et al. PLS path modeling [ J ].Computational Statistics and Data Analysis, 2005 (48) : 159 - 205.
[ 6 ] Wang C. M., Tong H. Q.. Best iterative initial values for PLS in aCSI model [J ]. Mathematical and Computer Modeling, 2007, 46 (3 - 4) : 439 -444.
[ 7 ] Tong H. Q.. Evaluation model and its iterative algorithm byalternating projection [ J ]. Mathematical and Computer Modeling, 1993, 18(8) : 55 - 60.
[ 8 ] 方开泰. 含有线性约束及非负回归系数的回归模型 [J ]. 计算数学 , 1985(7) : 97 - 102.
[ 9 ] 童恒庆, 熊丽, 彭慧. Self - organized path constraint neural networkstructure and algorithm [J ]. Neural In formation Proceeding, 2006, ( PartI): 457 - 466.
[ 10 ] 童恒庆. 理论计量经济学 [M ]. 北京 :科学出版社 , 2005: 12 - 86.
[ 11 ] 童恒庆. 数据分析与统计计算软件DASC [M /CD ]. 北京 :科学出版社,2005.
发明内容
A.发明基本步骤
以上叙述的单层或者多层结构方程模型都是针对一个对象建立的模型。一个国家或者一个行业有许多企业 (对象 )。如果每个企业都各自利用自己的样本建立模型,即使模型的两个方程结构都完全一样,但是由于样本数据不一样 ,得到的系数也不一样。这样的顾客满意度计算结果显然缺乏可比性。因此应该研究多对象的建模,既保留路径分析模型参数估计客观性的一面,又在各对象之间保持参数估计的统一性,使得计算结果具有更好的可比性。
假设有
Figure RE-833097DEST_PATH_IMAGE071
个对象需要测评,每个对象都是同样的结构方程,同样的
Figure RE-674014DEST_PATH_IMAGE020
个观测变量,都进行了
Figure RE-479159DEST_PATH_IMAGE021
次观测。对于每一个对象都可得到了一个
Figure RE-63724DEST_PATH_IMAGE022
观测数据块。将这些数据块纵向叠放形成一个
Figure RE-903504DEST_PATH_IMAGE198
矩阵
Figure RE-164590DEST_PATH_IMAGE046
。每一个对象都满足一个结构方程模型,如何将这些模型统一起来形成一个合理的模型群,本发明试图利用我们前期研究所提出的凸约束的广义线性回归模型,来统领这
Figure RE-722611DEST_PATH_IMAGE071
个结构方程模型。具体算法分 3个步骤进行。
(1)将多对象结构方程模型原始数据纵向叠放,利用基于配方约束的结构方程模型确定性算法统一求解。
Figure RE-845287DEST_PATH_IMAGE071
个对象看作是一个对象,对
Figure RE-805153DEST_PATH_IMAGE020
个观测变量进行
Figure RE-987873DEST_PATH_IMAGE199
次观测,得到
Figure RE-767610DEST_PATH_IMAGE198
矩阵
Figure RE-693978DEST_PATH_IMAGE046
。套用SEM模型和我们的确定性算法,得到结构方程模型中的系数
Figure RE-508350DEST_PATH_IMAGE200
Figure RE-612703DEST_PATH_IMAGE201
, 取
Figure RE-879736DEST_PATH_IMAGE202
,
Figure RE-609795DEST_PATH_IMAGE203
Figure RE-278674DEST_PATH_IMAGE204
Figure RE-803196DEST_PATH_IMAGE205
。于是
Figure RE-557525DEST_PATH_IMAGE012
个结构自变量分别有了权系数
Figure RE-91275DEST_PATH_IMAGE206
,
Figure RE-614660DEST_PATH_IMAGE009
个结构因变量分别有了权系数
Figure RE-559351DEST_PATH_IMAGE207
。此时的数据结构整体如图2,但是它的行数是
Figure RE-800977DEST_PATH_IMAGE198
,有
Figure RE-872838DEST_PATH_IMAGE071
个数据块纵向叠放。数据阵左侧部分如图10(多对象结构方程模型的数据排列图)。
这样求解得到每个结构变量对应的观测变量的汇总系数,为下一步使用评估模型提供系数约束条件。
(2) 将叠放的数据块按结构变量纵向剖分,分别采用评估模型求解,得到每个对象每个结构变量的评估分。
注意
Figure RE-250730DEST_PATH_IMAGE208
,
Figure RE-382634DEST_PATH_IMAGE209
是全体观测变量的个数,它分别从属于
Figure RE-845976DEST_PATH_IMAGE210
个结构变量。矩阵
Figure RE-721528DEST_PATH_IMAGE046
可以按列剖分成
Figure RE-953927DEST_PATH_IMAGE210
个数据块,称之为列数据块,每个列数据块对应一个结构变量
Figure RE-7464DEST_PATH_IMAGE023
或者
Figure RE-489261DEST_PATH_IMAGE027
。对于每个列数据块,每个列数据块的数据结构图都类似于图8,可套用前面叙述过的评估模型,即凸约束的广义线性回归模型,约束是
Figure RE-371767DEST_PATH_IMAGE211
以及
Figure RE-255409DEST_PATH_IMAGE212
或者
Figure RE-197957DEST_PATH_IMAGE213
,
Figure RE-167050DEST_PATH_IMAGE024
Figure RE-853246DEST_PATH_IMAGE028
是它的变量个数,评估对象都是
Figure RE-60237DEST_PATH_IMAGE071
个,
Figure RE-954112DEST_PATH_IMAGE214
或者
Figure RE-879343DEST_PATH_IMAGE215
是它的评估分,都是
Figure RE-634809DEST_PATH_IMAGE071
维列向量。这样就得到了每个结构变量下每个对象的评估分,形成了一个
Figure RE-961886DEST_PATH_IMAGE216
的矩阵
Figure RE-511816DEST_PATH_IMAGE217
。一共需要进行
Figure RE-924342DEST_PATH_IMAGE209
个评估模型的计算,每个评估模型都会得到
Figure RE-952341DEST_PATH_IMAGE071
个评估分。当然每个评估模型都需要进行一次独立完整的交互投影计算。
这样计算的结果相当于把原始数据压缩了,每个对象只剩下一行,这一行就是各个结构变量的评估分。
(3)将上一步计算得到的
Figure RE-665082DEST_PATH_IMAGE218
矩阵
Figure RE-402225DEST_PATH_IMAGE217
作为新的观测矩阵,按结构方程模型求解。
如同图2里的观测矩阵
Figure RE-567627DEST_PATH_IMAGE046
,代回到原来的结构方程模型。由于现在每个结构变量只对应一个观测变量,结构方程模型中的系数
Figure RE-133738DEST_PATH_IMAGE038
Figure RE-966565DEST_PATH_IMAGE039
或者
Figure RE-592718DEST_PATH_IMAGE200
Figure RE-245416DEST_PATH_IMAGE201
计算都是简单的。主要的计算任务是在结构方程式(2)中计算路径系数
Figure RE-615218DEST_PATH_IMAGE006
Figure RE-771393DEST_PATH_IMAGE219
。完成了结构方程的计算,顾客满意度所在的变量
Figure RE-348873DEST_PATH_IMAGE220
的估计值
Figure RE-957709DEST_PATH_IMAGE221
就计算出来了。
Figure RE-396781DEST_PATH_IMAGE221
Figure RE-407462DEST_PATH_IMAGE071
维向量,它的第
Figure RE-906577DEST_PATH_IMAGE222
个分量就是第
Figure RE-2709DEST_PATH_IMAGE222
个对象的顾客满意度的评估数值,
Figure RE-714313DEST_PATH_IMAGE223
这样计算的结果就得到每个对象的顾客满意度最终评估分。
B: 发明的关键技术。
(1)基于配方约束的结构方程模型确定性算法。
(2)基于凸集间交互投影的评估模型算法。
(3)基于配方约束和凸集间交互投影的多对象结构方程模型计算技术。
附图说明:
图1是一个中国顾客满意指数模型的变量与路径结构图。
图2是一个中国顾客满意指数模型的数据排列图。
图3是一个多层结构方程模型的变量与路径图。
图4是考试加总分的数据结构图。
图5是一元线性回归数据结构图。
图6是线性回归的最小二乘法则。
图7是线性回归最小二乘法则的投影几何意义。
图8是评估模型数据结构图。
图9是凸集间的交互投影算法示意图。
图10是多对象结构方程模型的数据排列图。

Claims (1)

1.本发明专利“基于凸集间交互投影的多对象结构方程模型计算技术”分三个步骤求解多对象结构方程模型。第一步,将多对象结构方程模型原始数据纵向叠放,利用基于配方约束的模型确定性算法统一求解,得到每个结构变量对应的观测变量的汇总系数。第二步,将叠放的数据块按结构变量纵向剖分,分别采用评估模型的凸集间的交互投影算法求解,得到每个对象每个结构变量的评估分。第三步,将上一步计算得到的评估分矩阵作为新的观测矩阵,按结构方程模型求解,得到每个对象的顾客满意度最终评估分。
CN202011067786.0A 2020-10-07 2020-10-07 基于凸集间交互投影的多对象结构方程模型计算技术 Pending CN112069450A (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011067786.0A CN112069450A (zh) 2020-10-07 2020-10-07 基于凸集间交互投影的多对象结构方程模型计算技术

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011067786.0A CN112069450A (zh) 2020-10-07 2020-10-07 基于凸集间交互投影的多对象结构方程模型计算技术

Publications (1)

Publication Number Publication Date
CN112069450A true CN112069450A (zh) 2020-12-11

Family

ID=73683657

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011067786.0A Pending CN112069450A (zh) 2020-10-07 2020-10-07 基于凸集间交互投影的多对象结构方程模型计算技术

Country Status (1)

Country Link
CN (1) CN112069450A (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106469177A (zh) * 2015-08-21 2017-03-01 中国传媒大学 基于结构方程模型的互联网电影用户满意度分析方法及系统
CN110322121A (zh) * 2019-06-12 2019-10-11 国网天津市电力公司 一种供电企业客户满意度评估方法
CN111125629A (zh) * 2019-12-25 2020-05-08 温州大学 一种域自适应的pls回归模型建模方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106469177A (zh) * 2015-08-21 2017-03-01 中国传媒大学 基于结构方程模型的互联网电影用户满意度分析方法及系统
CN110322121A (zh) * 2019-06-12 2019-10-11 国网天津市电力公司 一种供电企业客户满意度评估方法
CN111125629A (zh) * 2019-12-25 2020-05-08 温州大学 一种域自适应的pls回归模型建模方法

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
徐春艳;钟绍军;: "评估模型基于交互投影的参数岭估计", 湖北师范学院学报(自然科学版), no. 03, 26 September 2009 (2009-09-26) *
童乔凌 等: "结构方程模型的约束最小二乘解与确定性算法", 数值计算与计算机应用, vol. 30, no. 3, 30 September 2009 (2009-09-30), pages 1 - 5 *
童乔凌;邹雪城;熊丽;童恒庆;: "多层与多对象结构方程模型的约束最小二乘解", 武汉理工大学学报(信息与管理工程版), no. 06, 15 December 2009 (2009-12-15) *
童恒庆: "评估模型及其迭代解法", 应用数学, vol. 5, no. 2, 30 April 1992 (1992-04-30), pages 1 - 4 *
童恒庆;余超;赵旭杰;: "凸约束广义线性回归模型的参数估计及算法", 应用数学, no. 04, 15 October 2008 (2008-10-15) *

Similar Documents

Publication Publication Date Title
Myers et al. Response surface methodology: process and product optimization using designed experiments
Wei et al. Variable importance analysis: A comprehensive review
Sheng et al. Comprehensive monitoring of nonlinear processes based on concurrent kernel projection to latent structures
Santana et al. Influence analysis in skew-Birnbaum–Saunders regression models and applications
Fan et al. High-dimensional adaptive function-on-scalar regression
Kimaev et al. A comparison of efficient uncertainty quantification techniques for stochastic multiscale systems
Ferrie Quantum model averaging
Hong et al. Hybrid active learning method for non-probabilistic reliability analysis with multi-super-ellipsoidal model
Xu et al. Block average quantile regression for massive dataset
Li et al. Spectral stochastic isogeometric analysis for linear stability analysis of plate
Foldnes et al. On identification and non-normal simulation in ordinal covariance and item response models
Aouichaoui et al. Comparison of group-contribution and machine learning-based property prediction models with uncertainty quantification
Bi et al. Stochastic validation of structural FE-models based on hierarchical cluster analysis and advanced Monte Carlo simulation
Touzani et al. Smoothing spline analysis of variance approach for global sensitivity analysis of computer codes
Zhu et al. Mixture robust L1 probabilistic principal component regression and soft sensor application
CN112069450A (zh) 基于凸集间交互投影的多对象结构方程模型计算技术
Zong et al. Embedded software fault prediction based on back propagation neural network
Jin et al. Transfer learning with quantile regression
Wang et al. Estimation and model selection for nonparametric function-on-function regression
CN112949000B (zh) 基于卷积神经网络模型的构件残余应力反演方法
Zhang et al. Nonparametric quantile regression for homogeneity pursuit in panel data models
Wang et al. Improved empirical likelihood inference and variable selection for generalized linear models with longitudinal nonignorable dropouts
Lu Simultaneous inference for semiparametric mixed-effects joint models with skew distribution and covariate measurement error for longitudinal competing risks data analysis
Otis Software architecture for CALPHAD modeling of phase stability and transformations in alloy additive manufacturing processes
Rani et al. Generative adversarial wavelet neural operator: Application to fault detection and isolation of multivariate time series data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20230203

Address after: 430072 Hubei Province, Wuhan city Wuchang District of Wuhan University Luojiashan

Applicant after: WUHAN University

Address before: 430070 floor 10-south-4, building 10, Bowen garden, 210 Luonan street, Hongshan District, Wuhan City, Hubei Province

Applicant before: WUHAN ZHUXIN TECHNOLOGY Co.,Ltd.