Skip to main content

Table 4 Improvement in classification accuracy using majority voting ensembles. Optimal unweighted majority-voting ensemble classifiers were formed by selecting classifiers from all 8 classifiers for each feature set listed and the average classification accuracy for 10-fold cross-validation was calculated. A paired-t test was performed for each ensemble classifier against the previous neural network classifier for each feature subset (SLF15 and SLF16 were compared against the previous classifier for SLF8 and SLF13, respectively). Each ensemble classifier was also compared against the optimal classifier for each feature set listed in Table 2 (SLF15 and SLF16 were compared with the individual optimal classifiers for SLF8 and SLF13, respectively).

From: Boosting accuracy of automated classification of fluorescence microscope images for location proteomics

Feature Set Classifiers in the Optimal Majority-voting Ensemble Average classification accuracy (%) P-value of paired t test with previous results P-value of paired t test with optimal single classifier Classification Accuracy Upper Bound* (%)
SLF8 (2D) Exprbf-kernel SVM
AdaBoost
Bagging
89.4 0.003 0.08 95.5
SLF15 (2D) Rbf-kernel SVM
Exponential-rbf-kernel SVM
Polynomial-kernel SVM
91.5 0.0006 0.01 96.1
SLF13 (2D DNA) Rbf-kernel SVM
AdaBoost
Mixtures-of-Experts
90.7 0.003 0.03 95.6
SLF16 (2D DNA) Neural Network
Linear-kernel SVM
Exprbf-kernel SVM
Polynomial-kernel SVM
AdaBoost
92.3 0.003 0.02 96.6
SLF14 (3D) Neural Network
Linear-kernel SVM
Exprbf-kernel SVM
Polynomial-kernel SVM
AdaBoost
89.8 0.02 0.29 96.3
SLF10 (3D DNA) Linear-kernel SVM
Rbf-kernel SVM
Exprbf-kernel SVM
Mixtures-of-Experts
95.8 0.02 0.35 98.2
  1. * The upper bound of classification accuracy for a feature set is defined as the percentage of all images that could be correctly classified by at least one of the eight tested classifiers using that feature set.