Research On Multivariate Network Data Visualization

As the popularization of information technology and the widespread of computer theory, scientific, engineering, business as well as military areas have generated a large amount of multivariate network data. Researchers are in a pressing need of effective techniques and tools to understand and analyze the data. Nonetheless, information visualization, as an effective analysis tool for expressing abstract data, remains still in a comparatively naive level with respect to multivariate network data. Thus, it is not able to suffice analysts'requirement of exhibiting multivariate distributional and network associational features of the data, as well as the relation between them. Moreover, most existing multivariate network visualizations are not adaptive well to mixed data sets, which is a common issue for other information visualization approaches.Therefore, in light of the foregoing problems, this thesis proposes a new multivariate network visualization method named ASCNet, and investigates in depth the problems within, including the pre-process of mixed data sets, the placement of vertices and the lay-out of edges in multivariate networks. Thereafter, the effectiveness of ASCNet is demonstrated by a case study of the typical multivariate network data set– the data from international arms trading intelligence. In particular, the key research problems and novelty of this thesis include:Firstly, to propose a systematic framework of multivariate network visualization, and to design a new visualization paradigm based on the framework. Multivariate network visualization is still in the emerging stage, where lacks of a unified and mature method framework. On the basis of the information reference model proposed by Card, this thesis proposes a pertinent visualization framework for multivariate network, namely MulNetVis. It normalize the design of multivariate network visualization methods, based on which a direcly implementation is generated– ASCNet. This method comprises three main steps, specifically data transformation based mixed data pre-process, advanced star coordinates based vertices placement, and edge-merging and -routing based edge layout. These key techniques provide an underlying basis for the discussion of implementation of ASCNet.Secondly, to propose a general data processing technique for visualizing mixed data sets. Multivariate attributes to delineate objects often involve multiple data types, and they can be mapped into either numerical or categorical as per the semantic relation among variates. However, traditional information visualizations are only capable of dealing with one single data type. Thus, this thesis puts forward a data transformation technique, on the basis of correspondence analysis, to quantify categorical data into numerical, facilitating the visualization of mixed data sets containing both numerical and categorical data. During the transformation, regarding to the data sets with relatively large number of variate or with categorical data possessing high cardinality, a set of clustering based cardinality reduction strategies are designed to reduce the variate number in correspondence analysis, as well as to improve the computational efficiency. This technique fulfills the mixed data pre-processing of ASCNet.Thirdly, to propose an advanced star coordinates with variate configuration strategy. Classic star coordinates based multivariate visualizations have shortcomings such as disadvantage of expressing variate distributional information, and information loss resulted by dimensional reduction, onerous manual configuration of variate axes. To this end, this thesis presents the advance star coordinates. It utilizes diameter instead of radius as variate axis to clearly show distributional information; then, to reduce information loss, it maps data objects down to meaningful lower dimensional and visual space while trying to retain the consistency of the coordinates of multivariate objects in two spaces; further, it incorporates the variate axes configuration strategy by optimizing the order of axes and the included angles between, which simplifies the interactive process and reduces the time needed for further analysis. This technique fulfills the vertice placement of ASCNet.Fourthly, to propose an edge layout method based on edge-merging and–routing. ASCNet displays in advanced star coordinates according to multivariate attributes when there is no association between vertices. There would result in visual clutters because of excessive edge-crossing if we connected the vertices directly. Hence, this thesis suggests an edge layout algorithm on the basis of edge-merging and–routing, reducing the probability of edge-crossing by merging edges sourcing from the same vertex, and therefore, representing the association between settled vertices. This technique fulfills the edge layout of ASCNet.Lastly, to devise and implement a multivariate network prototype named MulNetVisPrototype. With this as a benchmark while meeting the needs of intelligence analysis, the international arms trading data is employed as a case study in order to demonstrate the usability of ASCNet and its critical techniques in a systematic and comprehensive way.In all, this thesis proposes a systematic and normalized multivariate network visualization framework, investigates a visualization method that enables expressing the multivariate and network features of multivariate network data as well as their relation, and implements a prototype that verifies the effectiveness and correctness of the discussed theory and methods. They strongly support users in analyzing and mining multivariate network data. Thus far, this thesis will not only push forward the development of multivariate network visualization, but also provide data analysts with a fresh pathway to fully mastering and deeply understating multivariate network data leveraging the proposed visualization approach.