Presented is a new tool for the display and analysis of N-dimensional data. This tool creates a powerful working environment that allows the strength and flexibility of the dimensional stacking technique to be utilized to its fullest potential.
After learning how to interpret the data mappings that the technique produces, a user should begin to develop a feel for the visual clues provided and be able to make informed decisions as to which mappings to examine. The ability to make simple selections of channels and speeds enables the novice user to change mappings rapidly and the more experienced user to easily select a desired mapping. Data driven view selection is a novel method for choosing views of potential interest. The shearing and rotation options facilitate detection and exploration of structure which is not orthogonal with any of the axes, while the clipping mechanism allows for easy isolation of regions of interest. Binning control, overlap control, color map manipulation, and N-D brushing all act as mechanisms to facilitate the viewing of data, and preprocessing operations on the data can help remove noise and enhance existing structure.
Like any implementation of a new technique, N-Land has some disadvantages. Most notably, since it allocates screen space for each discrete bin in the N-D space, it is very much affected by the ``curse of dimensionality'', resulting in mostly empty images for small data sets with high dimensionality. Also, as shown in Table 1, the number of channels to examine increases greatly with respect to the number of dimensions. This can make a manual browsing quite time consuming for large data sets of high dimensionality.
As this is a relatively new area of exploration, there are many exciting directions for future development. One possibility is to investigate methods for automatic fine tuning (based on texture analysis) of the images in a manner analogous to the manual shearing and clipping process now being employed. The result would be a semi-automatic projection pursuit system which uses human judgement to attempt to eliminate paths with low probability of interest. Other avenues of exploration include extending the classes of image operators and image analysis techniques modified to work in N-dimensional space. Some candidates might include morphological operations, pyramid-based processing, and model-based structure recognition tasks. Finally, alternate internal structures for the data should be studied to address efficiency issues, and the possibility of porting the system to a parallel architecture might be explored.