Dataset Concordia (targetcl. 4)

Basic characteristics Concordia (targetcl. 4)

400

target objects

The Concordia CENPARMI handwritten digits (16x16 pixels per image). Used in "Neural-Network Classifiers for Recognizing Totally Unconstrained Handwritten Numerals" by Cho, Sung-Bae. Download mat-file with Prtools dataset.

3600

outlier objects

256

features

Unsupervised PCA Concordia (targetcl. 4)

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 Concordia (targetcl. 4)

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

Results Concordia (targetcl. 4)

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

534, 0 outliers, AUC (x100) 5x strat. 10-fold
Classifiers Preproc
none unit var PCA 95\%
Gauss 91.5 (0.0) 89.2 (0.0) 93.5 (0.0)
Min.Cov.Determinant NaN (0.0) NaN (0.0) NaN (0.0)
Mixture of Gaussians 94.0 (1.2) 90.2 (1.1) 93.0 (1.1)
Naive Parzen 87.9 (0.0) 88.0 (0.0) 93.6 (0.0)
Parzen 93.6 (0.0) 89.7 (0.0) 91.6 (0.0)
k-means 91.0 (0.5) 87.2 (0.3) 86.7 (0.5)
1-Nearest Neighbors 93.6 (0.0) 89.6 (0.0) 91.6 (0.0)
k-Nearest Neighbors 93.6 (0.0) 89.6 (0.0) 91.6 (0.0)
Nearest-neighbor dist 90.4 (0.0) 80.7 (0.0) 84.9 (0.0)
Principal comp. 93.6 (0.0) 90.6 (0.0) 92.8 (0.0)
Self-Organ. Map 92.4 (0.2) 87.9 (0.1) 89.2 (0.4)
Auto-enc network NaN (0.0) NaN (0.0) 89.0 (0.0)
MST 93.6 (0.0) 89.6 (0.0) 92.1 (0.0)
L_1-ball 44.4 (0.0) 44.4 (0.0) 90.3 (0.0)
k-center 89.8 (1.4) 86.5 (0.4) 84.9 (2.0)
Support vector DD 93.3 (0.0) 47.1 (1.0) 91.5 (0.1)
Minimax Prob. DD 93.5 (0.0) 89.4 (0.0) 91.5 (0.0)
LinProg DD 91.3 (0.0) 88.4 (0.0) 85.9 (0.0)
Lof DD 94.5 (0.0) 91.0 (0.0) 91.6 (0.0)
Lof range DD 95.3 (0.0) 89.7 (0.0) 93.4 (0.0)
Loci DD 91.0 (0.0) 87.2 (0.0) 87.5 (0.0)

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