Databases Reference
In-Depth Information
4.1 Results Generated by Alternate Partitioning Algorithm
The HAC + P2 + FSR algorithm was executed on the same data sets with the
same parameters to obtain a comparison of the two partitioning techniques.
The results are shown in Table 3.
Note that although different cluster similarity measures were used in the
partitioning process, the average and minimum self-similarities in the follow-
ing table were using the average similarity measure. This allows for direct
comparison with the results given for HAC + P + FSR.
Note that although the maximum depth parameter for these runs was
set to six, just like the runs of the HAC+P+FSR algorithm, the use of the
alternate quality function naturally led the algorithm to produce a flatter,
bushier hierarchy without reaching the depth limit. As expected, the average
and minimum self-similarity for each cluster increased with depth in the tree.
However, comparison of Tables 2 and 3 show that the similarity measures
for the clusters produced by HAC+P2+FSR were better overall at each level
of the generated hierarchies, except for the first level. This is explained by
the anomalies introduced in the original runs with HAC + P + FSR by single-
query clusters. HAC + P2 + FSR generated more first-level clusters, so the
effect of single-query clusters was diluted in the average values.
HAC + P2 + FSR tends to create a larger number of clusters at each level,
making the tree shallower and bushier than HAC+P+FSR. In addition, the
implementation of HAC+P2+FSR takes significantly less CPU time during
the partitioning step.
Tabl e 3 .
Alternate clustering
Level
# clusters Avg. self-similarity Min. self-similarity
Federal Government Department
1
48
0.511
0.117
2
184
0.583
0.204
3
292
0.742
0.331
4
100
0.863
0.557
Consumer Electronics Manufacturer
1
48
0.441
0.064
2
234
0.603
0.221
3
300
0.810
0.333
4
72
0.879
0.518
Consumer Software Producer
1
48
0.502
0.131
2
179
0.654
0.251
3
248
0.755
0.375
4
159
0.850
0.417
5
31
0.918
0.589
