In R, I am using the randomForest and caret packages for a Random Forest (RF) regression task. Genarally, there are two options when fine-tuning a model:
- GridSearch,
- RandomSearch.
I have seen on various posts that GridSearch has limitations such as taht there are intangibles that cannot be addressed in a gridsearch, such as parameter interaction stabilizes at a slower rate than error so Bootstrap replicates should be adjusted not just on error but also based on number and complexity of parameters included in the model.
My questions are:
Examples online are welcomed.
Good afternoon,
I did not run these plots for the random forest, thus, i do not know if it is possible to generate that. However, a plot on the interaction of parameters can be generated for the Gradient Boosting Models. Please, see attached an example.
Ines
Sure. So by interactions we mean, for example, how the number of trees in random forest affects the R2 (or RMSE, or whatever)?
The principle of parameters tuning of your model. After preparing your data is to find the optimal parameters and splitting dataset, simultaneously with the cross-validation, the parameter tuning with a grid search (or not) is conducted. Each model technic has their own parameters. The interaction is just the combination of your parameters, for the GBM, it will be the number of tree, the height of tree, the shrinkage, the minimum of observations by terminal nodes or to create a new branch. With the package caret, you run all models with all possible combinations (this step takes a long time), for example for the graph that i published 6,375 were generated. Then, you obtain the best combination of parameters (i.e., the parameters to have the best rate of accuracy and the minimal rate of errors). Finally, you run your model with the optimised parameters and you assess it.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Probability
- Probability Theory
- Reliability