海量空间相互作用数据挖掘及可视化

发布时间：2018-03-14 11:35

本文选题：空间相互作用数据　切入点：起点终点流数据　出处：《华中科技大学》2016年博士论文　论文类型：学位论文

【摘要】：随着位置感知技术的发展和广泛应用,海量的空间数据变得越来越容易获取。空间相互作用(Spatial Interaction)数据,也被称为起点终点流(Origin-Destination Flow)数据,作为一种特殊的空间数据,描述两个地点之间的关联,或者两个地点之间的物体移动。该数据在日常生活中非常普遍。例如:人的移民,车辆的行径,动物的迁徙,还有疾病的传播等等。大量空间相互作用数据的研究有助于理解不同领域的复杂动态,包括城市规划,智能交通,人口学,应急管理等等。例如,研究人类移动带来的位置到位置的关联程度,可以发现伴随而来的病毒传播的规律。目前的分析方法很难从大量的空间相互作用数据中发现有用的信息,也很难完全可视化大规模的数据,这是数据本身的复杂性所决定的。数据本身的复杂性主要包括:(1)大的数据量:一个中型规模的数据集通常包含几百到几千个位置,很容易有几千或者几百万位置与位置之间的关联。(2)多种数据维度:这些数据集一般镶嵌在多种维度内,比如空间维度,时间维度,网络维度,多属性维度。(3)可更改区域单元问题(Modifiable Area Unit Problem,MAUP):数据的起始位置的形状和大小差异非常大,导致数据集的分析产生偏差。(4)多尺度问题:数据在不同的地理或时间尺度上展示不同的规律。对于这样复杂的数据集,单一的方法不可能完全的解决这些问题。本文提出了一组方法来分析和可视化大规模的空间相互作用数据,并有效地处理了上述的问题。这些方法主要包括:基于共享邻居数的(Shared Nearest Neighbors)空间点聚类方法[1],该方法按照数据点分布的规律,将数据点归纳成含有相似数据量的聚类。在点分布密集的区域,形成小的地理空间的上的聚类,在点分布稀疏的区域,形成大的地理空间上的聚类。同时,这是一种数据驱动的聚类方法,能够在数据之间找到自然分割。在聚类的结果的基础上可视化统计量度,从而发现时空模式。该方法有效地处理了上述问题中数据量和可更改区域单元问题。基于共享邻居数的空间相互作用数据层次聚类方法[2],该方法将传统的层次聚类方法扩展运用到空间相互作用数据。该方法的主要思想是,将空间相互作用数据的起点间的相似度和终点间的相似度统一为一个相似性量度,然后运用聚合式的层次聚类方法将空间相互作用数据进行聚类。本文创新地使用共享邻居数作为相似性量度,来应对数据在空间上分布不均匀的问题。文章在对空间相互作用数据聚类的基础上,进一步研究了数据类群的时间特征。该方法能有效地处理上述问题中的前三个问题。空间相互作用数据核密度估计模型[3],这个方法的主要思想是将空间点的核密度估计方法扩展运用到空间相互作用数据,对相互作用数据集进行密度的估计,然后从密度分布中提取特征。该方法将数据进行了高度的抽象,能够很大程度上减少数据量。同时,核密度估计模型使用自适应的带宽(Adaptive Bandwidth),可以很好的处理数据分布不均的问题(可更改区域单元问题)。基于空间相互作用数据核密度估计模型的多尺度可视化,该方法主要将空间相互作用数据核密度估计模型进一步扩展。使用不同的参数进行空间相互作用数据密度估计和代表性数据选择,创新地实现了多辨率多尺度的流向地图。该方法能够非常有效的处理多尺度的问题。这些方法看似独立,但是它们从不同的角度分析了空间相互作用数据,用不同的方法增强了对复杂数据的理解。本文将多种不同但是互相补充的方法综合到一起,将计算方法、可视化方法、可视化分析方法等结合在一起分析空间相互作用数据,试图将不同方法的不同角度综合,形成一个全局整体的理解。
[Abstract]:With the development of Location Aware Technology and wide application of spatial data becomes more and more easy to get. The spatial interaction (Spatial Interaction), also known as the starting point and end point flow (Origin-Destination Flow) data, as a special kind of spatial data, describe the association between the two locations, moving objects or between two place. The data is very common in our daily life. For example: immigration, vehicle behavior, animal migration, and the spread of the disease and so on. A lot of research of spatial interaction data are useful in understanding the complex dynamic in different fields, including city planning, intelligent transportation, demography, emergency management and so on. For example correlation study of human, from a moving position to the position, can be found along with the spread of the virus to the law. The current analytical methods is very difficult from a large number of spatial interaction number According to the discovery of useful information, it is difficult to complete visualization of large-scale data, which is decided by the complexity of the data itself. The complexity of the data itself mainly includes: (1) a large amount of data: a medium-sized data sets often contain hundreds to thousands of locations, easy connection between thousands or millions of the position and position. (2) multiple data dimensions: These data sets usually embedded in multiple dimensions, such as spatial dimension, time dimension, network dimension, multi attribute dimension. (3) modifiable areal unit problem (Modifiable Area Unit Problem, MAUP) data: the starting position of the shape and size of the difference is very large the data set, analysis causes the deviation. (4) Multiscale Problems: data show different patterns in different geographical or time scales. For such complex data sets, a single method can not completely solve these The problem. This paper presents a set of methods to spatial analysis and visualization of large-scale interaction data, and effectively handle the problems. These methods include: Based on the number of shared neighbors (Shared Nearest Neighbors) [1] spatial clustering method, the method according to the data distribution, data points are summed up with clustering similar amount of data. In the distribution of dense area clustering form small geographical space on the distribution in sparse regions, cluster large geographical space. At the same time, this is a clustering method for data driven, can be found in the data between the natural segmentation based on clustering results. The visualization of statistical measure, in order to find the spatiotemporal pattern. This method effectively solves the problem of data and can change the unit number of shared neighbors. Spatial interaction based on data Hierarchical clustering method of [2], the traditional hierarchical clustering method is applied to the expansion of space interaction data. The main idea of this method is that the similarity of unified spatial interaction between the starting point and end point data similarity between a similarity measure, and then use the aggregation hierarchical clustering method of space interaction data type clustering. This paper using the number of shared neighbors as the similarity measure to deal with the data in a spatially inhomogeneous problem. Based on the spatial interaction data clustering, time to study the features of data groups. This method can effectively deal with the first three problems above problems in space. The interaction data of kernel density estimation model of [3], the main idea of this method is that the space point kernel density estimation method is extended to the spatial interaction of data The interaction data set of density estimation, and then extract features from density distribution. The data were highly abstracted, can greatly reduce the amount of data. At the same time, kernel density estimation using adaptive bandwidth model (Adaptive Bandwidth), can the problem of uneven distribution of data well (change the regional unit problem). Multiscale visualization of spatial interaction data model based on kernel density estimation method, the main spatial interaction data kernel density estimation model is further extended. Using different parameters of spatial interaction data density estimation and representative data selection, innovation to achieve a multi resolution multi-scale flow map. This method can deal with multi scale problem is very effective. These methods seem to be independent, but they are from a different point of view of the spatial interaction data, Using different methods to enhance the understanding of complex data. This will be different but mutually complementary methods together, the calculation method, visualization, visualization and analysis methods combined with analysis of spatial interaction data, to different methods with different angles to form a comprehensive, overall understanding.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：P208

【相似文献】