Skip to main content

Advertisement

Table 1 Cross validated (32 runs of 3:2 train-test split) linear SVM and random forest performance on original datasets, RST transformed data sets, LDA transformed data sets and PCA transformed data sets

From: A subspace recursive and selective feature transformation method for classification tasks

   Original rst lda pca
  No. features Log loss ± stdev Log loss ± stdev Log loss ± stdev Log loss ± stdev
Otto group
 RandomForest 93 (original) 1.60 ± 0.03    
  80   2.12 ± 0.04   1.69 ± 0.03
  68   3.44 ± 0.08   1.71 ± 0.03
  58   3.39 ± 0.08   1.69 ± 0.02
  49   3.36 ± 0.08   1.74 ± 0.02
  41   3.36 ± 0.15   1.72 ± 0.03
  34   3.25 ± 0.18   1.74 ± 0.03
  8    2.12 ± 0.06  
 LinearSVM 93 (original) 0.96 ± 0.04    
  80   0.87 ± 0.04   0.78 ± 0.01
  68   0.77 ± 0.05   0.79 ± 0.03
  58   0.70 ± 0.01   0.80 ± 0.03
  49   0.70 ± 0.01   0.81 ± 0.02
  41   0.70 ± 0.01   0.84 ± 0.03
  34   0.70 ± 0.01   0.88 ± 0.03
  8    0.94 ± 0.02  
mnist
 RandomForest 784 (original) 0.46 ± 0.00    
  684   1.24 ± 0.04   1.55 ± 0.04
  597   1.86 ± 0.04   1.50 ± 0.04
  454   1.85 ± 0.05   1.53 ± 0.07
  396   1.87 ± 0.03   1.46 ± 0.03
  345   1.98 ± 0.14   1.33 ± 0.02
  288   1.26 ± 0.20   1.29 ± 0.03
  9    1.04 ± 0.02  
 LinearSVM 784 (original) 1.85 ± 0.10    
  684   1.17 ± 0.06   1.33 ± 0.05
  597   0.60 ± 0.09   1.33 ± 0.04
  521   0.68 ± 0.13   1.37 ± 0.11
  396   0.71 ± 0.07   1.35 ± 0.07
  345   0.55 ± 0.09   1.30 ± 0.06
  288   0.41 ± 0.11   1.37 ± 0.07
  9    0.52 ± 0.01  
Olivetti faces
 RandomForest 4096 (original) 4.29 ± 1.09    
  720   6.45 ± 0.58   9.58 ± 1.43
  509   3.67 ± 0.40   8.24 ± 0.38
  430   3.32 ± 0.53   8.04 ± 0.75
  375   3.62 ± 0.39   8.13 ± 1.05
  327   4.42 ± 0.34   7.29 ± 0.52
  288   3.96 ± 0.83   7.05 ± 0.83
  39    4.15 ± 0.26  
 LinearSVM 4096 (original) 1.98 ± 0.04    
  720   1.21 ± 0.05   1.51 ± 0.09
  509   1.07 ± 0.04   1.50 ± 0.04
  430   1.07 ± 0.04   1.53 ± 0.08
  375   1.07 ± 0.04   1.51 ± 0.05
  327   1.07 ± 0.04   1.52 ± 0.09
  285   1.07 ± 0.04   1.52 ± 0.06
  39    1.53 ± 0.07  
  1. In bold are the best score (lowest log loss) on the corresponding data set. Results for otto group classification, mnist digits recognition and olivetti faces recognition