In contrast, each iteration of boosted MP-EST requires much less time: 12 minutes per iteration for SSG-boosting and 7 minutes per iteration for DACTAL-boosting, each run sequentially.
The three tuning parameters are the depth of the tree (even values from 2 to 10 were evaluated), the number of boosting iterations (20 iterations to 2000 in 100 iteration increments) and the learning rate (fixed at 0.1).
Thus, when used with MRP as a starting tree, the entire pipeline (computing the starting tree, running five iterations of DACTAL boosting, and selecting the final tree) completes in 35 minutes.
The number of trees (i.e., the number of boosting iterations) and interaction depth (i.e., the depth of trees) are the two parameters tuned.
The gbm.perf function estimates the optimal number of boosting iterations for a gbm object and optionally plots various performance measures (Ridgeway 2017).
The number of boosting iterations was selected using cross-validation.
The algorithms include: support vector machine with linear kernel (SVM/Poly, trained by sequential minimization optimization algorithm), SVM with RBF kernels (SVM/RBF), alternating decision tree (ADTree with number of boosting iterations set to 10), and k-nearest neighbor classifier (IBk with inverse distance weighing where k was determined by leave-one-out cross-validation).
The best and most parsimonious classifier was an alternating decision tree on H3K327me3 and H3K36me3 gene body signal, with default parameters (number of boosting iterations 10, searchpath = 'expand all paths').
On the other hand, the best quartet support score was nearly always found in a subsequent iteration, for both types of boosting techniques.
The best weak classifier is hence selected in each iteration of the boosting procedure such that the cost of (11) is the lowest as follows: (14).
