Parameter Tuning Presets for crossValidate and Their Customisation
Dario Strbenac
The University of Sydney, Australia.
presets.RmdWhen a choice about a model to use is made and specified to
crossValidate, performance tuning of a range of typical
values is automatically made. These values can be changed or removed.
Below is a table of models and their default values, which can be
changed by specifying the extraParams parameter. It is a
list with named elements. Each element must be named one of
"select", "train" or "predict"
(as well as "prepare" for data preparation which is not a
parameter tuning operation, so is not discussed here), to identify which
modelling the parameter corresponds to.
| Model | Keyword | Default Parameters | Other Parameters |
|---|---|---|---|
| Random Forest | "randomForest" |
mTryProportion = c(0.25, 0.33, 0.50, 0.66, 0.75, 1.00)
and num.trees = c(10, seq(100, 500, 100)
|
See ?ranger::ranger
|
| Random Survival Forest | "randomSurvivalForest" |
mTryProportion = c(0.25, 0.33, 0.50, 0.66, 0.75, 1.00)
and num.trees = c(10, seq(100, 500, 100)
|
See ?randomForestSRC::rfsrc
|
| Extreme Gradient Boosting | "XGB" |
mTryProportion = c(0.25, 0.33, 0.50, 0.66, 0.75, 1.00)
and nrounds = c(5, 10, 15)
|
See ?xgboost::xgboost
|
| k Nearest Neighbours | "kNN" |
k = 1:5 |
See ?class::knn
|
| Support Vector Machine | "SVM" |
kernel = c("linear", "polynomial", "radial", "sigmoid")
and cost = 10^(-3:3))
|
See ?e1071::svm
|
| naive Bayes Kernel | "naiveBayes" |
difference = c("unweighted", "weighted") |
weighting = c("height difference", "crossover distance")
and minDifference, a number 0 or higher. |
| Mixtures of Normals | "mixturesNormals" |
nbCluster = 1:2 |
difference = c("unweighted", "weighted")
and
weighting = c("height difference", "crossover distance")
and minDifference, a number 0 or higher. |
Note that the last two models are not wrappers but custom
implementations in this package. The weighting parameter specifies how
the differences between classes are calculated. Either the height
difference of a class’s fitted density to another class’s nearest
density is used or the horizontal distance from the measurement value to
the nearest position where the density of that class crosses any other
class. The distances are only used if difference is
weighted. minDifference specifies the minimum
distance between densities for a feature to have a vote for a particular
class. The default of 0 means that every feature votes for the class
which has the highest density at that point, regardless how close the
density of any other class is to it.