Dataset Glass building float

Basic characteristics Glass building float

70

target objects

The Glass Identification database from UCI, originally to distinguish 6 types of glass by their oxide content. The building_windows_float_processed class is used as target class. Download mat-file with Prtools dataset.

144

outlier objects

9

features

Unsupervised PCA Glass building float

On the left, the PCA scatterplot is shown, on the right the retained variance for varying number of features.
On the left, the PCA scatterplot is shown of data rescaled to unit variance, on the right the retained variance.

Supervised Fisher Glass building float

On the left, the Fisher scatterplot is shown, on the right the ROC curve along this direction.

Results Glass building float

The experiments are performed using dd_tools. A rudimentary explanation of the classifiers is given in the classifier section.

582, 0 outliers, AUC (x100) 5x strat. 10-fold
Classifiers Preproc
none unit var PCA 95\%
Gauss 78.6 ( 1.0) 79.7 ( 0.6) 77.7 ( 1.6)
Min.Cov.Determinant NaN ( 0.0) NaN ( 0.0) 79.7 ( 1.3)
Mixture of Gaussians 82.6 ( 1.0) 83.4 ( 1.3) 80.7 ( 1.3)
Naive Parzen 76.1 ( 0.8) 76.1 ( 0.8) 79.4 ( 1.5)
Parzen 85.5 ( 1.0) 83.8 ( 0.8) 81.3 ( 1.3)
k-means 79.2 ( 1.1) 81.2 ( 0.7) 78.6 ( 1.0)
1-Nearest Neighbors 85.9 ( 1.1) 85.2 ( 0.9) 84.6 ( 1.3)
k-Nearest Neighbors 85.9 ( 1.1) 85.2 ( 0.9) 84.6 ( 1.3)
knn, opt-AUC 85.2 ( 1.3) 83.8 ( 1.1) 83.4 ( 0.4)
Nearest-neighbor dist 81.4 ( 1.6) 82.3 ( 0.9) 79.7 ( 2.3)
Principal comp. 78.4 ( 1.4) 80.7 ( 0.8) 79.4 ( 1.0)
Self-Organ. Map 83.9 ( 1.4) 84.9 ( 1.3) 83.2 ( 1.9)
Auto-enc network 76.7 ( 1.0) 81.3 ( 1.2) 78.1 ( 2.4)
Spanning Tree 86.1 ( 1.0) 86.0 ( 0.9) 85.1 ( 0.7)
L_1-ball 59.3 ( 1.7) 73.3 ( 1.1) 80.1 ( 1.4)
k-center 82.8 ( 1.8) 81.9 ( 1.2) 83.0 ( 2.0)
Support vector DD 75.9 ( 3.4) 83.7 ( 1.1) 75.9 ( 3.2)
Minimax Prob. DD 79.2 ( 1.5) 84.8 ( 0.6) 75.7 ( 1.3)
LinProg DD 75.3 ( 1.5) 81.6 ( 1.2) 74.9 ( 1.2)

Classifier projection spaces The first classifier projection spaces are obtained by computing the classifier label disagreements (setting the threshold on 10% target error) and applying an MDS on the resulting distance matrix between classifiers:



Original



Unit variance



PCA mapped

Classifier projection spaces The second versions of the classifier projection spaces are obtained by computing the classifier ranking disagreements and applying an MDS on the resulting distance matrix between classifiers:



Original



Unit variance



PCA mapped