Dataset Sonar (targetcl. mines)

Basic characteristics Sonar (targetcl. mines)

111

target objects

The Sonar dataset from the undocumented databases from UCI. The task is to train a network to discriminate between sonar signals bounced off a metal cylinder and those bounced off a roughly cylindrical rock. Download mat-file with Prtools dataset.

97

outlier objects

60

features

Unsupervised PCA Sonar (targetcl. mines)

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 Sonar (targetcl. mines)

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

Results Sonar (targetcl. mines)

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

508, 0 outliers, AUC (x100) 5x strat. 10-fold
Classifiers Preproc
none unit var PCA 95\%
Gauss 70.0 (0.8) 65.7 (0.8) 69.7 (0.5)
Min.Cov.Determinant NaN (0.0) NaN (0.0) 67.7 (1.0)
Mixture of Gaussians 76.5 (0.8) 64.3 (1.5) 79.2 (0.9)
Naive Parzen 56.9 (1.1) 56.9 (1.2) 69.9 (0.8)
Parzen 84.8 (0.6) 69.5 (0.8) 85.3 (0.8)
k-means 72.4 (1.3) 62.5 (1.6) 70.0 (1.2)
1-Nearest Neighbors 84.8 (0.7) 69.8 (0.6) 86.2 (0.8)
k-Nearest Neighbors 84.8 (0.7) 69.8 (0.6) 86.2 (0.8)
Nearest-neighbor dist 77.8 (1.0) 84.9 (0.7) 77.4 (1.2)
Principal comp. 73.1 (1.0) 60.8 (0.9) 68.3 (1.8)
Self-Organ. Map 84.0 (1.0) 71.1 (1.2) 84.2 (1.3)
Auto-enc network 71.5 (0.8) 59.4 (1.4) 68.7 (1.4)
MST 84.9 (0.7) 71.5 (0.6) 86.6 (0.4)
L_1-ball 39.0 (2.5) 54.0 (1.9) 65.5 (0.4)
k-center 78.5 (2.4) 62.2 (1.3) 76.1 (1.8)
Support vector DD 75.8 (0.9) 70.5 (5.4) 71.9 (1.5)
Minimax Prob. DD 83.7 (0.3) 69.6 (0.8) 82.7 (1.0)
LinProg DD 57.1 (1.1) 64.4 (0.5) 50.7 (1.1)
Lof DD 86.8 (1.5) 84.3 (0.8) 84.8 (0.6)
Lof range DD 85.5 (1.2) 78.7 (0.8) 83.9 (1.2)
Loci DD 70.7 (1.5) 67.2 (0.8) 70.8 (1.9)

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