Performance of the SVMs were tested using a three-way cross-validated experiment. The gene expression vectors were randomly divided into three groups. Classifiers were trained using two thirds of the data and tested on the remaining third. This procedure was then repeated two more times, each time using a different third of the genes as test genes.
Four separate SVM classifiers were tested, each employing a different kernel function. Three of the four kernel functions used were based on a dot-product (mercer) kernel, differing by the power to which the kernel was raised (1, 2, or 3). The fourth used a radial basis function kernel.
Functional classes defined by the MIPS Yeast Genome Database were used to train and test SVMs and include the following six classes: tricarboxylic acid cycle (TCA) (targetmips~0), respiration (targetmips~1), cytoplasmic ribosomes (targetmips~2), proteasome (targetmips~3), histones (targetmips~4) and helix-turn-helix proteins (targetmips~5) .
The performance of each classifier was measured by examining how well the classifier identified the positive and negative examples in the test sets. Each gene in the test set can be categorized in one of four ways: true positives (TP) are class members according to both the classifier and MYGD; true negatives (TN) are non-members according to both; false positives (FP) are genes that the classifier places within the given class, but MYGD classifies as non-members; false negatives (FN) are genes that the classifier places outside the class, but MYGD classifies as members. We report the number of genes in each of these four categories for each of the learning methods we tested.
To judge overall performance, we define the cost of using the method M as C(M) = fp(M) + 2 * fn(M) , where fp(M) is the number of false positives for method M, and fn(M) is the number of false negatives for method M. The false negatives are weighted more heavily than the false positives because, for these data, the number of positive examples is small compared to the number of negatives. The cost for each method is compared to the cost C(N) for using the null learning procedure, which classifies all test examples as negative. We define the cost savings of using the learning procedure M as S(M) = C(N) - C(M) .