Average Performance of Support Vector Machines (SVMs) With a Radial Kernel and Multilayer Perceptrons (MLPs) With 79 Input Units, Eight Hidden Units, and Five Output Units When Trained to Learn the Gene Expression Profiles of Different Functional Classes
| Class | Class size | Method | False positive | False negative | True positive | True Negative |
| TCA cycle | 17 | MLP | 11.0 ± 3 | 10.0 ± 2 | 7.0 ± 2 | 2439.0 ± 3 |
| SVM | 5.6 | 9.0 | 8.0 | 2444.4 | ||
| Respiratory processes | 27 | MLP | 6.8 ± 1 | 9.6 ± 1 | 17.4 ± 1 | 2433.2 ± 1 |
| 30 | SVM | 6.0 | 10.4 | 19.6 | 2431.0 | |
| Ribosomal genes | 121 | MLP | 9.4 ± 3 | 4.0 ± 2 | 117.0 ± 2 | 2336.6 ± 3 |
| SVM | 5.4 | 5.4 | 115.6 | 2340.6 | ||
| Proteasome | 35 | MLP | 11.6 ± 2 | 8.6 ± 1 | 26.4 ± 1 | 2420.4 ± 2 |
| SVM | 1.8 | 7.0 | 28.0 | 2430.2 | ||
| Histone related genes | 11 | MLP | 0.6 ± 1 | 2.0 ± 1 | 9.0 ± 1 | 2455.4 ± 1 |
| SVM | 0.0 | 2.0 | 9.0 | 2456.0 |
-
The averages are taken over five independent realizations of the three-fold cross-validation scheme. For MLP, the standard deviations over the five different realizations is included. False negatives show similar performance between the two methods, but MLP tend to produce a higher number of false positives. Of the 30 respiratory genes used to train the SVM, 3 were also members of the TCA cycle, and were excluded from the respiratory process set when training the MLP.
