Information Technology Reference
In-Depth Information
Fig. 16.5.
Indices ability to identify the optimal clustering solution - application on
DS1 and DS2 for document clustering
Fig. 16.6.
Indices ability to identify the optimal clustering solution - application on
DS3 and DS4 for word clustering
16.4.6
Discussion
After examining the different graphs in Figures 16.5, 16.6, 16.7, and 16.8, we
can conclude the following remarks:
On involving indices as external indicators.
An important first remark is
that most indices perform “well” at evaluating solutions, especially when applied
on words. By looking at Figures 16.5, 16.6, we can notice the high correlations
that most indices have with the
FScore
, which means that they have compara-
ble behaviors to an external index relying on a predefined structure. For word
clustering, on the top of the list, we can find the
C
2andthe
H
3 indices with
the outstanding correlations of 0.936 and 0.938 on DS3 and DS4 respectively,
which are of course encouraging results. For document clustering, correlations
were less satisfying. Nevertheless, we could obtain very high correlations with
three indices, i.e.,
H
3,
C
3,
C
4, but very low correlations, which could barely go
beyond 0.2, were obtained with the other indices.
