Research On Visual Analytics For Multidimensional Data And Micro-blogging Social Network

Visual analytics is a multi-disciplinary subject, which plays an increasingly im-portant role in understanding large-scale data, mining and network data analysis, etc.Through efective visualization techniques, users can participate in the process of knowl-edge discovery and data mining. This thesis focus on applying the visual analysis tech-niques to multidimensional data, data clustering analysis and the currently popular socialnetworking, to reveal the trends and rules in the datasets.In multidimensional data analysis, we propose a new visual pattern detection tech-nique to reveal the implied relations inside multidimensional datasets. Although the tra-ditional multidimensional data analysis techniques, such as Parallel Coordinates, can re-veal the overall patterns of the data, the visual results is directly related to the axis ordermapped by data dimensions, and it is also not easy to observe the corresponding changesof values and patterns caused by the value changing on one specifc dimension. Thisthesis carries out data processing to obtain the correlation information and relatively thebest pairwise combination of dimensions in two levels-data level and visual presentationlevel, and extract the patterns to guide data reorder. The proposed layout is generated bya series of automated calculation and interactive adjustment. Through mapping the datavalues, patterns and gradient information to three concentric circle groups, respectivelyand several interactive features to support users to globally or locally rearrange data andhighlight patterns, the corresponding patterns, gradients and the values on other dimen-sions will be changed and re-distributed. Therefore, the inherent patterns will be present.Finally, the validity of this method is verifed by applying several experimental datasets.Cluster analysis is an important knowledge discovery technique. However, thereis no universal clustering algorithm which can be applied to all application felds. Thisthesis designs and implements a multidimensional data clustering analysis frameworkin a visual interactive environment. In the cluster generation stage, multidimensionaldatasets are mapped to node-link diagram by using Force-Directed algorithms to revealits inner structure. Then the graph clustering based on Markov flow algorithm is used tocluster the visualizing results adjusted by the user. Then, we map the multidimensionaldataset to Sammon plane and support the user to manually select clustering centers. Inaddition, we also integrate k-means algorithm in the framework. In the cluster evaluation stage, by visual analysis of the generated parallel clustering view, we can detect the coredata records, adjust dimension weights, and remove noise data to improve the clusteringperformance. Meanwhile, we support users’ feedback to train the data clustering modelbetter. This method can efectively solve the problems of clustering algorithms, suchas the selection of number of clusters, weight parameters of dimensions and clusteringuncertainty.Finally, we investigate how information propagation in a specifc microbloggingplatform evolves and carry out visual mining tasks to detect how information circu-lates, identify relevant patterns, understand dynamic attributes of information propaga-tion and the underlying sociological motivations. We propose three efcient strategies:a Hierarchy Aware Force-Directed layout, an accumulated fltering algorithm and micro-aggregation layout to map the multiple data attributes to appropriate visual elements,highlighting those major key elements while maintaining an optimized and elegant lay-out and revealing the inter/intra-connections between diferent micro-groups. For thedynamic attributes, we propose two models: the depth-varying parallel data model andthe time-varying parallel data model to illustrate the temporal evolution efciently. Wealso develop several interactive features to manipulate and navigate the retweeting prop-agation graph in convenient ways. We demonstrate how our approaches increase the un-derstanding of the retweeting propagation graph from a visual perspective by employingretweeting propagation datasets collected from Sina Weibo. Meanwhile, this visual min-ing tool has been evaluated by data analysts and successfully used in Weibo departmentof Sina Corporation as a helpful assistant to them.