基于近邻方法的高维数据可视化聚类发现

VISUAL CLUSTERING FOR HIGH DIMENSIONAL DATA BASED ON NEAREST NEIGHBOR

提出了一种新颖的基于近邻方法的高维数据可视化聚类方法，并实现了一个近邻可视化聚类发现系统VisNN.已有的解决高维数据可视化聚类方法主要是通过降维把高维数据（高维属性空间中的点，每个属性代表一个维度）投影到二维或三维空间上，从而达到可视化目的．在此采取了一种完全不同的处理方法：构成可视平面的2个坐标轴不再是降纸后的两个属性或属性组合，而是记录（一个记录即高维属性空间中的一个点）序列．通过对记录按用户关心的属性进行排序，两个记录在高维空间中的相似关系（表示为距离）在由以记录为坐标轴构成的平面中仍然有可能得到保持．实验结果表明，用户通过使用该系统能较为容易地发现聚类和异常现象．

A new visual clustering method for high dimensional data is presented, and a general visualization system VisNN is implemented in this paper. Classical visual method for clustering high dimension data is to project high dimensional data into two-dimension or three-dimension space by using dimension reduction method. In the method for each axis of two-dimension X-Y space, a record series, instead of a reduced attribute, are used. When the order of records is sorted by some attributes that users concern about, the distance relationship of two records in high dimensional space could be kept to some extent in the X-Y space. The experiments show that VisNN can help users discover interesting clusters and abnormality easily.