Data Mining - Mehmed Kantardzic [261]
For example, visualization techniques can be used for efficient process of “visual clustering.” The algorithm is based on finding a set of projections P = [P1, P2, … ,Pk] useful for separating the initial data into clusters. Each projection represents the histogram information of the point density in the projected space. The most important information about a projection is whether it contains well-separated clusters. Note that well-separated clusters in one projection could result from more than one cluster in the original space. Figure 15.10 shows an illustration of these projections. You can see that the axes’ parallel projections do not preserve well the information necessary for clustering. Additional projections A and B, in Figure 15.10, define three clusters in the initial data set.
Figure 15.10. An example of the need for general projections, which are not parallel to axes, to improve clustering process.
Visual techniques that preserve some characteristics of the data set can be invaluable for obtaining good separators in a clustering process. In contrast to dimension-reduction approaches such as PCAs, this visual approach does not require that a single projection preserve all clusters. In the projections, some clusters may overlap and therefore not be distinguishable, such as projection A in Figure 15.10. The algorithm only needs projections that separate the data set into at least two subsets without dividing any clusters. The subsets may then be refined using other projections and possibly partitioned further based on separators in other projections. Based on the visual representation of the projections, it is possible to find clusters with unexpected characteristics (shapes, dependencies) that would be very difficult or impossible to find by tuning the parameter settings of automatic-clustering algorithms.
In general, model visualization and exploratory data analysis (EDA) are data-mining tasks in which visualization techniques have played a major role. Model visualization is the process of using visual techniques to make the discovered knowledge understandable and interpretable by humans. Techniques range from simple scatter plots and histograms to sophisticated multidimensional visualizations and animations. These visualization techniques are being used not only to convey mining results more understandable to end users, but also to help them understand how the algorithm works. EDA, on the other hand, is the interactive exploration of usually graphical representations of a data set without heavy dependence on preconceived assumptions and models, thus attempting to identify interesting and previously unknown patterns. Visual data-exploration techniques are designed to take advantage of the powerful visual capabilities of human beings. They can support users in formulating hypotheses about the data that may be useful in further stages of the mining process.
15.7 REVIEW QUESTIONS AND PROBLEMS
1. Explain the power of n-dimensional visualization as a data-mining technique. What are the phases of data mining supported by data visualization?
2. What are fundamental experiences in human perception we would build into effective visualization tools?
3. Discuss the differences between scientific visualization and information visualization.
4. The following is the data set X:
Although the following visualization techniques are not explained with enough details in this book, use your knowledge from earlier studies of statistics and