Information Technology Reference
In-Depth Information
Table 2. Observed patterns obtained from 2D and 3D cluster plots.
category
Db
CI
a*
b*
Energy Entropy
HP
medium
low
narrow
mid-range
low
medium
mid-range
P
high
mid-range
high
HC
low
low
narrow
narrow
mid-range
mid-range
C
low
very high
high
H
low Db
or high CI
Fig. 6. 2D cluster plot of Db-CI.
have adverse effects on the cluster viewing (in other words, the precision was
higher than what was required). We therefore tried different quantization scales
and finally chose a 10-interval scale which was the coarsest scale that could still
distinguish each data point. The plots for each triplet of features (with a* and b*
being always kept together) re-confirm the observations summarized in Table 2.
We also wished to determine from the 3D cluster plots which triplets of
features give better discriminating power, in order to use these as a starting
point to gradually explore rules for clustering in higher dimensions using parallel
coordinates. Figure 7 shows a 3D cluster plot (Db, a*, b*) for Healthy and Polyps
cases. However, we observed that although there appeared to be some groupings,
these groupings were not separable and the rules underlying these groupings were
not simple. We investigated this problem further using another visualization
method based on parallel coordinates, which allowed the simultaneous viewing
of multi (more than 3) dimensions.
6.2
Discovery of Rules Using Parallel Coordinates Plots
High dimensional data is often transformed or projected to 2D or 3D representa-
tions for visualization. However, this practice usually causes a loss of information.
Parallel coordinates allow n -dimensional data to be displayed in 2D [51]. In this
method, n Cartesian coordinates are mapped into n parallel coordinates, and an
Search WWH ::




Custom Search