基于马田系统的动态多维贫困精准识别方法研究

Accurate recognition is the premise and foundation of taking targeted measures in poverty alleviation, and also the key to assess and consolidate the achievements of taking targeted measures in poverty alleviation. With the advent of the big data era, the governments from central to local level are actively exploring to take advantage of big data techniques to accurately recognize the poor population. However, the explorations of the related techniques and methods are relatively lagged in academic circles. Faced with this situation, this research proposes an accurate recognition method based on Mahalanobis Taguchi System from the dynamic and multidimensional perspectives and the method can be applied to big data platforms. The research framework of the method is based on Mahalanobis Taguchi System. The accurate recognition is the research goal of the method. The method mainly deals with five aspects: measurement table, measurement data, measurement scale, and measure benchmark and recognition thresholds. The main contents of this research include: the participatory poverty evaluation method is introduced to design measurement indexes; the method of standardized heterogeneous data is put forward; the kernel function theory is applied to construct multiple kernel mahalanobis distance; three dimensional data is unfolded into two dimensional data by multi-way statistical analysis technology; nonparametric method is proposed to calculate the multi-threshold based on a single measurement benchmark. This research can help improve the working mechanism of taking targeted measures in poverty alleviation, optimize the resource allocation efficiency for accurate poverty alleviation, improve the performance of taking targeted measures in poverty alleviation, consolidate the achievements of taking targeted measures in poverty alleviation, and promote the integration of big data and poverty alleviation.

精准识别是精准扶贫的前提和基础，也是考核精准扶贫成效、巩固精准扶贫成果的关键。随着大数据时代的到来，从中央到地方都在积极探索利用大数据技术解决贫困人口的精准识别问题，然而学术界对相关的技术和方法探索却相对滞后。针对这种情况，本课题引入马田系统理论，从动态、多维视角提出一种能够应用于大数据平台的贫困精准识别方法。该方法以马田系统理论为研究框架，以精准识别为目标导向，从贫困的测量表、测量数据、测量尺度、测量基准和识别阈值等五个方面展开研究。主要研究内容包括：采用参与式贫困评估法设计测量表；提出异构数据的同构化方法；利用核函数理论构建多核马氏距离；利用多向分析技术将三维数据展开成二维数据进行统计研究；提出单基准多识别阈值的划定方法。通过本课题的研究，可以进一步完善精准扶贫工作机制，优化扶贫资源配置效率，提升精准扶贫成效，巩固精准扶贫成果，促进大数据与扶贫治理加速融合。

探索利用大数据技术，从动态多维视角精准识别贫困，不仅是精准扶贫的前提条件，也是2020年后治理相对贫困的重要保障。本课题以马田系统理论为研究框架，从贫困的“测量表、测量数据、测量尺度、测量基准和识别阈值”等五个方面，构建了完整的动态多维贫困精准识别理论和方法。课题取得了以下研究成果。（1）基于可持续生计框架，设计了两套动态多维贫困测量表。（2）提出采用多向统计分析和符号分析技术，将动态多维贫困数据打包压缩成符号区间数据，并进而构建了一种能够处理符号区间数据的马田系统。（3）针对动态多维贫困数据往往具有非线性结构的问题，首先利用核技巧直接推导了一种核马氏距离，然后针对动态多维贫困数据具有高噪性、高维性、非线性和不平衡性等特点，将马田系统和核主成分两种方法相融合，构建了一种核主成分马田系统。为验证两种方法的有效性，分别采用随机模拟数据和实地调研数据进行验证，结果表明两种方法均优于传统马田系统，能够对动态多维贫困进行精准识别。（4）针对传统马田系统只能在原始特征空间中进行测量基准优化的问题，课题将其拓展到由核函数诱导的特征空间，给出了相应的正交优化规则；（5）为揭示致贫的深层次原因，课题将马田系统分别与解释结构模型和社会网络分析法相结合，构建了两种致贫深层次原因分析方法。