Visualization of knowledge structures: some methodological issues

摘要

There is a growing interest in the application of mapping techniques to gain insights into the complex structures of scientific and technological developments. Cluster analysis and projection techniques (like multidimensional scaling, principal components analysis, and self organizing maps) are essential components of the toolkits that are used for mapping of science. Hierarchical and non-hierarchical clustering techniques (e.g. k-means) assume that boundaries between clusters (i.e., domains or sub-domains of science) are sharp and do not overlap. How far is this assumption correct? Further, the projection of high-dimensional data onto low-dimensional spaces (usually two dimensional plots) inevitably leads to distortions, and greater the dimensionality of the data, greater is the distortion in the projection of individual objects. Here, a crucial question is: What is the extent of distortion and how can we assess it? This paper responds to these concerns and illustrates the application of fuzzy cluster analysis and a technique (SPTNNE) for visualization of projection distortions.