Dataset Glass containers

Basic characteristics Glass containers

13

target objects

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

201

outlier objects

9

features

Unsupervised PCA Glass containers

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 containers

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

Results Glass containers

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

585, 0 outliers, AUC (x100) 5x strat. 10-fold
Classifiers Preproc
none unit var PCA 95\%
Gauss 91.0 ( 2.3) 88.1 ( 2.1) 74.8 ( 6.2)
Min.Cov.Determinant NaN ( 0.0) NaN ( 0.0) 76.4 ( 3.1)
Mixture of Gaussians 90.8 ( 2.4) 82.8 ( 3.4) 88.3 ( 2.9)
Naive Parzen 77.9 ( 4.6) 77.9 ( 4.6) 83.2 ( 3.3)
Parzen 90.1 ( 2.3) 82.7 ( 2.8) 74.9 ( 6.2)
k-means 83.6 ( 7.0) 79.5 ( 2.4) 77.4 ( 7.1)
1-Nearest Neighbors 91.0 ( 2.2) 82.1 ( 3.1) 89.8 ( 1.7)
k-Nearest Neighbors 91.0 ( 2.2) 82.1 ( 3.1) 89.8 ( 1.7)
knn, opt-AUC NaN ( 0.0) NaN ( 0.0) NaN ( 0.0)
Nearest-neighbor dist 90.5 ( 1.4) 79.5 ( 2.4) 89.7 ( 1.6)
Principal comp. 89.4 ( 2.3) 90.5 ( 2.0) 68.4 ( 4.3)
Self-Organ. Map 91.1 ( 2.2) 82.5 ( 3.0) 90.4 ( 1.4)
Auto-enc network 80.9 ( 5.4) 80.5 ( 5.0) 90.8 ( 3.8)
Spanning Tree 91.2 ( 2.2) 82.0 ( 3.0) 92.4 ( 1.1)
L_1-ball 38.4 ( 5.4) 43.1 ( 3.7) 81.7 ( 4.3)
k-center NaN ( 0.0) NaN ( 0.0) NaN ( 0.0)
Support vector DD 89.7 ( 0.3) 82.8 ( 2.4) 83.3 (11.9)
Minimax Prob. DD 91.0 ( 2.2) 82.2 ( 3.1) 90.7 ( 2.1)
LinProg DD 91.1 ( 2.2) 82.2 ( 2.8) 90.5 ( 2.1)

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