Author: Zhu Li
Date: 2019/10/05
Code:
clusters_for_digits.py -> clustering algorithms for the digit dataset and their performance evaluation
figures_for_digits.py -> drawing to display for the result
GaussianHelper.py -> a helper for GaussianMixture(@Zhao Yue)
Output for clusters_for_digit.py：
=========== ========================================================
date: 2019/09/25
author: zhu li
=========== ========================================================
Shorthand full name
--------------------------------------------------------------------
nor Normalized Mutual Information
homo Homogeneity
comp Completeness
--------------------------------------------------------------------
ten_means MiniBatchKMeans
AffinityPro AffinityPropagation
MeanShift MeanShift
Spectral SpectralClustering
Ward Ward
Agglom AgglomerativeClustering
DBSCAN DBSCAN
=========== ========================================================
n_digits: 10, n_samples 1797, n_features 64
__________________________________________________________________________________
init time inertia homo compl v-meas ARI- AMI- silhouette
k-means++ 0.17s 69676 0.683 0.722 0.702 0.572 0.699 0.140
random 0.12s 69415 0.604 0.652 0.627 0.467 0.623 0.147
PCA-based 0.03s 70769 0.669 0.696 0.683 0.558 0.679 0.147
__________________________________________________________________________________
__________________________________________________________________________________
clusters time inertia nor- homo comp
MB_KMeans 0.02s 3034 0.444 0.437 0.452
AffinityPro 6.44s 0 0.458 0.476 0.441
MeanShift 0.77s 0 0.445 0.460 0.431
Spectral 0.22s 0 0.461 0.459 0.462
Ward 0.08s 0 0.451 0.444 0.458
Agglom 0.08s 0 0.053 0.010 0.279
DBSCAN 0.02s 0 0.308 0.316 0.301
GaussianMix 0.03s 0 0.470 0.463 0.477
__________________________________________________________________________________
output for figures_of_digits.py:
Code:
cluster_document.py -> clustering algorithms for the document dataset and their performance evaluation
为了加速，实验时先用了PCA和LDA(Latent Dirichlet Allocation)降维，将向量化之后的数据降到低维（2，3，4，5维）再运行算法。
print(doc)
=========== ========================================================
date: 2019/10/05
author: zhu li
=========== ========================================================
Shorthand full name
--------------------------------------------------------------------
nor Normalized Mutual Information
homo Homogeneity
comp Completeness
--------------------------------------------------------------------
ten_means MiniBatchKMeans
AffinityPro AffinityPropagation
MeanShift MeanShift
Spectral SpectralClustering
Ward Ward
Agglom AgglomerativeClustering
DBSCAN DBSCAN
=========== ========================================================
output:
PCA(n_components=2):
__________________________________________________________________________________
clusters time inertia norm homo comp
MB_KMeans 0.05s 4 0.422 0.486 0.366
AffinityPro 6.08s 0 0.411 1.000 0.169
MeanShift 0.68s 0 0.236 0.228 0.245
Spectral 0.43s 0 0.430 0.539 0.344
Ward 0.16s 0 0.401 0.447 0.361
Agglom 0.14s 0 0.026 0.005 0.147
DBSCAN 0.15s 0 0.000 0.000 1.000
GaussianMix 0.06s 0 0.431 0.490 0.378
__________________________________________________________________________________
PCA(n_components=3):
__________________________________________________________________________________
clusters time inertia norm homo comp
MB_KMeans 0.02s 8 0.467 0.559 0.389
AffinityPro 5.81s 0 0.411 1.000 0.169
MeanShift 2.16s 0 0.240 0.215 0.269
Spectral 0.36s 0 0.499 0.604 0.412
Ward 0.18s 0 0.498 0.592 0.419
Agglom 0.24s 0 0.022 0.004 0.134
DBSCAN 0.17s 0 0.000 0.000 1.000
GaussianMix 0.09s 0 0.462 0.561 0.380
__________________________________________________________________________________
PCA(n_components=4):
__________________________________________________________________________________
clusters time inertia norm homo comp
MB_KMeans 0.04s 14 0.452 0.544 0.375
AffinityPro 6.08s 0 0.411 1.000 0.169
MeanShift 1.53s 0 0.190 0.160 0.225
Spectral 0.41s 0 0.517 0.620 0.431
Ward 0.19s 0 0.489 0.577 0.416
Agglom 0.34s 0 0.028 0.005 0.169
DBSCAN 0.20s 0 0.000 0.000 1.000
GaussianMix 0.09s 0 0.434 0.519 0.362
__________________________________________________________________________________
PCA(n_components=5):
__________________________________________________________________________________
clusters time inertia norm homo comp
MB_KMeans 0.02s 18 0.456 0.538 0.387
AffinityPro 6.11s 0 0.411 1.000 0.169
MeanShift 1.46s 0 0.233 0.229 0.238
Spectral 0.42s 0 0.507 0.597 0.431
Ward 0.22s 0 0.452 0.521 0.391
Agglom 0.40s 0 0.030 0.005 0.170
DBSCAN 0.21s 0 0.000 0.000 1.000
GaussianMix 0.14s 0 0.433 0.528 0.354
__________________________________________________________________________________
LatentDirichletAllocation(n_components=2):
__________________________________________________________________________________
clusters time inertia norm homo comp
MB_KMeans 0.03s 2 0.381 0.447 0.325
AffinityPro 6.17s 0 0.411 1.000 0.169
MeanShift 0.27s 0 0.417 0.449 0.386
Spectral 1.84s 0 0.337 0.435 0.261
Ward 0.15s 0 0.382 0.447 0.326
Agglom 0.11s 0 0.516 0.429 0.620
DBSCAN 0.06s 0 0.000 0.000 1.000
GaussianMix 0.03s 0 0.391 0.448 0.342
__________________________________________________________________________________
致谢：
感谢赵跃同学写的GuassianHelper.py，提供了便利。让高斯混合模型可以直接和别的模型一起处理，不用单独处理了。