PMC:4996413 / 26950-28433 JSONTXT

{"target":"https://pubannotation.org/docs/sourcedb/PMC/sourceid/4996413","sourcedb":"PMC","sourceid":"4996413","source_url":"https://www.ncbi.nlm.nih.gov/pmc/4996413","text":"2.4.3. Classification\nFollowing the selection of CpG sites (used as features from now on), classification is performed to construct classifiers fed by the selected features and measure their performance, thus validating the relevance of the selected features. Three nearest-neighbor classifiers (k-nn, k = 1, 6, 12) with weights, a classification tree (the GINI index was used as a split criterion), and a feed-forward artificial neural network (ANN) of one hidden layer were used. All classification algorithms, except ANN, are described in more detail in [28]. The ANN used here was trained using the back-propagation algorithm for 1000 epochs with a learning rate equal to 0.3 and momentum equal to 0.2 that were found to be the best choices on a trial and error basis. The hidden layer used a sigmoid activation function and contained ((num. of features+num. of classes)/2 + 1) nodes. Classifiers’ performance, in terms of total accuracy (number of samples correctly classified), and class sensitivity (number of true positives in a class that were correctly classified in this class) was measured using a training set and leave-one-out resampling (tabular results presented in Supplementary Material). The same classification algorithms were evaluated, utilizing the totally unknown-independent testing set: classifiers were constructed once using the training set and applied to the samples belonging to the testing set (results presented here in tabular format and bar plots).","divisions":[{"label":"Title","span":{"begin":0,"end":21}}],"tracks":[]}