基于维度扩展的Radviz可视化聚类分析方法

Extending Dimensions in Radviz for Visual Clustering Analysis

Radviz是一种多维数据可视化技术,它通过径向投影机制将多维数据映射到低维空间,使具有相似特征的数据点投影到相近位置,从而形成可视化聚类效果.Radviz圆周上的维度排列顺序对数据投影结果影响很大,提出将原始维度划分为多个新维度来拓展Radviz圆周上的维度排序空间,从而获得比原始维度条件下更好的可视化聚类效果.该维度划分方法首先计算数据在每个原始维度的概率分布直方图,然后使用均值漂移算法对直方图进行划分,最后根据划分结果将原始维度扩展为多个新维度.提出使用Dunn指数和正确率来量化评估Radviz可视化聚类效果.进行了多组对比实验,结果表明,维度扩展有利于多维数据在Radviz投影中获得更好的可视化聚类效果.

Radviz is a radial visualization technique that maps data from multi-dimensional space onto a planar picture. The dimensions placed on the circumference of a circle, called dimension anchors, can be reordered to reveal different patterns in the dataset. Extending the number of dimensions can enhance the flexibility in the placement of dimension anchors to explore meaningful visualizations. This paper describes a method that rationally extends a dimension to multiple new dimensions in Radviz. This method first calculates the probability distribution histogram of a dimension. The mean shift algorithm is applied to get centers of probability density to segment the histogram, and then the dimension can be extended according to the number of segments of the histogram. The paper also suggests using Dunn＇s index and accuracy rate to find the optimal placement of DAs, so the better effect of visual clustering can be achieved and evaluated after the dimension expansion in Radviz. Finally, it demonstrates the effectiveness of the new approach on synthetic and real world datasets.