Dataset Glass building nonfloat

Basic characteristics Glass building nonfloat

76

target objects

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

138

outlier objects

9

features

Unsupervised PCA Glass building nonfloat

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 nonfloat

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

Results Glass building nonfloat

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

583, 0 outliers, AUC (x100) 5x strat. 10-fold
Classifiers Preproc
none unit var PCA 95\%
Gauss 66.0 ( 0.9) 65.5 ( 0.6) 61.6 ( 1.2)
Min.Cov.Determinant NaN ( 0.0) NaN ( 0.0) 63.6 ( 0.9)
Mixture of Gaussians 67.5 ( 1.0) 67.9 ( 0.8) 67.3 ( 0.9)
Naive Parzen 64.1 ( 0.6) 64.1 ( 0.6) 68.7 ( 1.3)
Parzen 72.0 ( 0.9) 68.3 ( 0.7) 68.9 ( 1.1)
k-means 70.0 ( 1.7) 68.9 ( 2.4) 69.6 ( 1.9)
1-Nearest Neighbors 74.8 ( 1.6) 72.2 ( 1.4) 74.4 ( 1.9)
k-Nearest Neighbors 74.8 ( 1.6) 72.2 ( 1.4) 74.4 ( 1.9)
knn, opt-AUC 74.7 ( 1.9) 71.7 ( 1.3) 73.5 ( 2.4)
Nearest-neighbor dist 71.0 ( 2.1) 63.4 ( 2.0) 75.7 ( 2.1)
Principal comp. 67.9 ( 0.8) 66.1 ( 1.3) 63.3 ( 1.6)
Self-Organ. Map 70.8 ( 1.3) 69.6 ( 1.8) 70.4 ( 2.1)
Auto-enc network 65.9 ( 3.1) 66.1 ( 2.1) 64.8 ( 1.2)
Spanning Tree 74.9 ( 1.0) 72.5 ( 1.1) 74.1 ( 2.3)
L_1-ball 47.7 ( 0.8) 58.7 ( 1.2) 60.4 ( 2.0)
k-center 71.2 ( 0.5) 70.4 ( 1.0) 71.7 ( 2.3)
Support vector DD 70.1 ( 2.1) 69.1 ( 1.9) 69.8 ( 2.0)
Minimax Prob. DD 69.5 ( 0.8) 68.2 ( 0.9) 70.1 ( 2.4)
LinProg DD 60.7 ( 1.5) 69.3 ( 0.6) 58.9 ( 2.3)

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