Skip to main content

Table 2 Performance of oligo kernel classifiers on an enlarged data set with four times the number of examples than utilized for the EcoGene-based analysis with results shown in table 1. The oligomer length again varies according to K = 1,...,6. The table shows the mean classification error, given in percent, on the test sets. The rates are averages over 20 runs on randomly partitioned data with the same proportions of training, validation and test sets as for the previous results shown in table 1. According to the main paradigm of machine learning we would expect the error to decrease for an increased data set. However, obviously this is not the case, as the error rates are rising up to 6.4 percent, as compared with table 1. Therefore the results indicate that the additional data which have not been experimentally verified, are distributed in a different way than the verified TIS sequences from EcoGene. For that reason we conclude that these additional data should not be used for analysis of TIS, because it cannot be excluded that the distinct distribution is due to erroneous annotation.

From: Oligo kernels for datamining on biological sequences: a case study on prokaryotic translation initiation sites

oligomer length 1 2 3 4 5 6
mean error 17.3 15.6 15.3 16.0 17.0 18.9