BN-Mapping:基于贝叶斯网络的地理空间数据可视分析

BN-Mapping:Visual Analysis of Geospatial Data with Bayesian Network

地理空间数据指的是带有显式地理位置信息的数据,此类数据通常是多源、多变量的,其各变量之间的因果关系是位置相关的、异构的及多面性的,这给多变量地理空间数据的分析带来了很大的挑战.文中基于贝叶斯网络,针对地理空间数据提出了一种新颖的可视分析方法.首先,文中用一种启发式的方法对原始的多变量地理空间数据进行离散化,将连续的变量转换成离散的变量;然后,该方法对离散化之后的数据进行训练,得到相应的贝叶斯网络模型;接着,该方法针对贝叶斯网络模型的拓扑结构、条件概率表及数据本身基于地理位置的分布关系等分别设计了可视化视图并进行可视化;最后,文中设计了一个可交互的可视分析界面,将多个可视化视图进行了有机的结合,允许用户对变量之间的地理位置相关的隐式关系进行交互的探索、发现及分析.文中对多个数据集进行了包括异常检测、地理分类和因果关系分析在内的多种任务的案例分析,结果表明文中提出的可视分析方法可以十分有效地帮助用户对地理空间数据进行分析.

Geospatial data is those data with explicit geographic position information.Usually,it is multi-source as well as multivariate.Analyzing multivariate geospatial data is challenging because the causal relationships among multiple variates are location-dependent,heterogeneous and multi-faceted.In this paper,we propose a novel visual analysis approach for geospatial data based on Bayesian Network （BN）.Firstly,we propose a heuristic discretion method,which discretizes the original continuous geospatial data into a discrete one;Then,a Bayesian Network model is learned from the discretized multivariate geospatial dataset;After that,we designed three visual representations,namely,a global causal relationship in the form of a graph,detailed relationship quantities and probability in the form of a pixel-map,and a salient BN-distribution in the form of pixel charts,for the topology and conditional probability table of the learned Bayesian Network, and the geospatial distribution,respectively.Finally,we designed an interactive visual analysis interface,which well combines all the views and allows users to interactively explore,discover and reason about implicit patterns in a location-aware fashion.Case studies on several datasets for various analysis tasks,including anomaly detection,location-aware classification,and causal relationship discovery,demonstrate the efficiency of our approach in analyzing geospatial data.