Skip to main content

Table 3 Comparison of classification accuracy using top ranked features for platform transition

From: Evaluation of data discretization methods to derive platform independent isoform expression signatures for multi-class tumor subtyping

Feature selection CV (#) SVM-RFE (#) RF_based_FS (#)
Classifier FC Equal-W Equal-F k-means FC Equal-W Equal-F k-means FC Equal-W Equal-F k-means
SVM 43.4 (500) 35.5 (80) 100 (700) 92.1 (300) 51.3 (400) 75.0 (200) 100 (1000) 73.6 (60) 48.6 (20) 39.4 (50) 97.3 (600) 92.1 (200)
RF 69.7 (300) 84.2 (1000) 97.3 (1000) 89.4 (600) 61.8 (60) 89.4 (700) 96.0 (1000) 81.5 (100) 73.6 (40) 85.5 (100) 97.3 (800) 88.1 (300)
NB 27.6 (800) 30.2 (10) 92.1 (500) 75 (200) 35.5 (40) 38.1 (10) 85.5 (600) 67.1 (60) 35.5 (200) 34.2 (20) 94.7 (600) 78.9 (90)
PAM 44.7 (300) 26.3 (10) 92.1 (400) 76.3 (300) 44.7 (900) 39.4 (600) 89.4 (400) 60.5 (60) 46.0 (10) 34.2 (10) 93.4 (500) 82.8 (200)
  1. # Number of variables in the classification model
  2. Comparison of classification methods trained on exon-array (342 samples) and tested on RNA-seq (76 samples). The best accuracy (percentage of samples correctly predicted) achieved by each combination of the four classifiers and three feature selection schemes are presented, with number of features used in the best fitted model is shown in parenthesis. The models were built by stepwise addition of feature variables into the model by considering the top 1,000 ranked feature variables. Highest accuracy, achieved with the least number of features, for each classification method is marked in bold.