ARMOPSO Lab Report
In data clustering, we choose 5 artificial datasets and 9 UCI datasets to test the performance of ARMOPSO. Table 4 shows the specific information of these datasets.
Table 4 The information of datasets used in data clustering.
Datasets Number of data objects Dimension of a data object Number of clusters
Art2
data3 data1 Art1 data9 Iris
Wine
Glass
Thyroid
Ecoli bupaLD Cancer pimaIndians CMC 250
400
400
600
600
150
178
214
215
336
345
683
768
1473 3
2
2
2
3
4
13
9
5
7
6
9
8
9 5
3
2
4
3
3
3
6
3
8
2
2
2
3
3.2.2 Parameter setting
During the experiment, the parameter settings for each algorithm are presented in Table 5. In Tales 6, 9 and10, best means the best results in 20 independent runs. And mean represents the average results in 20 independent …show more content…
From data in the table we can get the following information. In best experimental results: for experimental results ARMOPSO generated, only on two datasets, that is, Glass and CMC, ARMOPSO does not produce better clustering results compared with other algorithms. ARMOPSO achieves better clustering results on the remaining 12 datasets. Especially on datasets data1 and Ecoli, the improvement of NMI is obvious. In mean experimental results, ARMOPSO achieves better clustering results on 7 datasets, compared with other algorithms. And there is also a dataset whose NMI equals to that in DSDE, and they both achieve the best clustering result. This fully illustrates that ARMOPSO has a great advantage for data clustering compared with other …show more content…
Therefore, we first extract image feature of each face image. We cluster ORL and Yale after obtaining the low dimensional image vectors. In this paper, we use principal component analysis (PCA) to extract eigenface feature [35] and obtain eigenspace X firstly. Then, eigenvectors in X are arranged in descending order based on their corresponding eigenvalues. Since the vast majority information of people’s face is included in the top 5% -10% eignvectors in descending order, we only choose the top 10% of eigenvetors to compose eignspace . Finally, each face image is projected into eignspace to obtain a new image vector. The new image vector achieves dimension reduction effect and retains most of the original image’s
Table 4 The information of datasets used in data clustering.
Datasets Number of data objects Dimension of a data object Number of clusters
Art2
data3 data1 Art1 data9 Iris
Wine
Glass
Thyroid
Ecoli bupaLD Cancer pimaIndians CMC 250
400
400
600
600
150
178
214
215
336
345
683
768
1473 3
2
2
2
3
4
13
9
5
7
6
9
8
9 5
3
2
4
3
3
3
6
3
8
2
2
2
3
3.2.2 Parameter setting
During the experiment, the parameter settings for each algorithm are presented in Table 5. In Tales 6, 9 and10, best means the best results in 20 independent runs. And mean represents the average results in 20 independent …show more content…
From data in the table we can get the following information. In best experimental results: for experimental results ARMOPSO generated, only on two datasets, that is, Glass and CMC, ARMOPSO does not produce better clustering results compared with other algorithms. ARMOPSO achieves better clustering results on the remaining 12 datasets. Especially on datasets data1 and Ecoli, the improvement of NMI is obvious. In mean experimental results, ARMOPSO achieves better clustering results on 7 datasets, compared with other algorithms. And there is also a dataset whose NMI equals to that in DSDE, and they both achieve the best clustering result. This fully illustrates that ARMOPSO has a great advantage for data clustering compared with other …show more content…
Therefore, we first extract image feature of each face image. We cluster ORL and Yale after obtaining the low dimensional image vectors. In this paper, we use principal component analysis (PCA) to extract eigenface feature [35] and obtain eigenspace X firstly. Then, eigenvectors in X are arranged in descending order based on their corresponding eigenvalues. Since the vast majority information of people’s face is included in the top 5% -10% eignvectors in descending order, we only choose the top 10% of eigenvetors to compose eignspace . Finally, each face image is projected into eignspace to obtain a new image vector. The new image vector achieves dimension reduction effect and retains most of the original image’s