A Projected Clustering Algorithm and Its Biomedical Application - Clustering Challenges in Biological Network - page 201

tered incorrectly. Therefore, we define the accuracy as the percentage of points
that are correctly clustered by the algorithm. For the typical result, the accuracy
of IPROCLUS is 91.45% and that of PROCLUS is 70.46%. Both algorithms are
tested multiple times. On average, IPROCLUS and PROCLUS can achieve the
accuracy of 91.90% and 70.18% respectively. In a word, IPROCLUS can achieve
much better accuracy than PROCLUS for the scaled dataset in the random case.

Table 9.3. Confusion Matrix for IPROCLUS and PROCLUS

IPROCLUS

Input

A B C D E Outliers

Output
1 3 0 97 22893 39 272

2 1525 1 60 120 428 668

3 0 23854 162 0 463 555

4 1 4 25031 15 348 583

5 14800 0 311 65 422 393

Outliers 439 0 17 0 3902 2529

PROCLUS

Input

A B C D E Outliers

Output
1

16626 0 2 406 366 260

2 2 22794 12 1346 461 356

3 6 1 2565 6953 1221 1098

4 3 0 19116 0 0 11

5 131 361 310 4803 1450 1181

Outliers 0 703 3673 9585 2104 2094

Then we consider the accuracy for the unscaled dataset in the random case.
The data used are generated in exactly the same way as in PROCLUS paper. The
two algorithms have compatible results. The average accuracy of IPROCLUS and
PROCLUS are 93.02% and 93.94% respectively.
Table 9.4 gives the average accuracy result for the two extreme cases. Similar
to the random case, IPROCLUS exhibits much better accuracy than PROCLUS
for the scaled datasets and compatible accuracy for the unscaled datasets.

Table 9.4. The Accuracy Result for the Two Extreme Cases

Data PROCLUS IPROCLUS

All-same-dimensions scaled 75.27% 96%

All-same-dimensions unscaled 95.91% 94.57%

No-common-dimensions scaled 51.08% 87.75%

No-common-dimensions unscaled 87.67% 86.87%

In summary, IPROCLUS can achieve much better accuracy than PROCLUS
for the scaled data in all the three cases. When the unscaled data is considered,