How to retrieve optimal MinLeafSize after automatic hyperparameter optimization for Tree Ensemble (fitrensemble)?

Hi. I am running MATLAB's automatic Bayesian optimization for a number of parameters for a Tree Ensemble.

opts = struct('Kfold',4,'Optimizer','bayesopt');
Mdl = fitrensemble(X,Y,'OptimizeHyperparameters',{'Method','NumLearningCycles','LearnRate','MinLeafSize'},'HyperparameterOptimizationOptions',opts);

I understand that all the optimal parameters are embedded in the resulted object ‘Mdl’, but I was wondering if it’s possible to retrieve and save in a variable the optimal MinLeafSize. Even though I have found the rest optimized parameters:

Mdl.ModelParameters.Method %Method
Mdl.ModelParameters.NLearn %NumLearningCycles
Mdl.ModelParameters.LearnRate %LearnRate

but, I cannot obtain the MinLeafSize. However, I can see that it is listed among the properties of 'Mdl' under MinLeaf:

Mdl.ModelParameters.LearnerTemplates{1,1}

Anyone knows how to extract this? Thanks.

Answer 1

[Cris LaPierre]

I ran both a tree and ensemble models optimizing minLeafSize. For a decision tree, MinLeaf is a model parameter, but not for an ensemble. The only way I could find to see the value was by viewing the template.

Mdl.ModelParameters.LearnerTemplates{1,1}

ans = 

Fit template for regression Tree.

        SplitCriterion: []
             MinParent: []
               MinLeaf: 126
             MaxSplits: 10
          NVarToSample: []
           MergeLeaves: 'off'
                 Prune: 'off'
        PruneCriterion: []
               QEToler: []
            NSurrogate: []
                MaxCat: []
                AlgCat: []
    PredictorSelection: []
          UseChisqTest: []
                Stream: []
          Reproducible: []
               Version: 2
                Method: 'Tree'
                  Type: 'regression'

[Haris K.]

Hi Cris, thank you for your answer.

Well, one could as well argue that it is a paramter of fitrensemble(). You can pass it indirectly via the Learner Name-Vlaue pair.

tr = templateTree('MinLeafSize',1,'MinParentSize',10);
opts = struct('Kfold',4,'Optimizer','bayesopt');
Mdl = fitrensemble(X,Y,'Learners',tr,'OptimizeHyperparameters',{'Method','NumLearningCycles','LearnRate'},'HyperparameterOptimizationOptions',opts);

So essentially, what you have written in your answer is what I also found myself. I guess I need someone to confirm if we cannot actually save the optimal parameter in a variable.

[Bernhard Suhm]

You can use bestPoint(Mdl.HyperparameterOptimizationResults) to access the hyperparameters of the "best estimated" model, including 'MinLeafSize'

[Haris K.]

Perfect! Exactly what I was looking. It might be worth adding this somewhere in the examples or in the inputs section of fitrtree(), fitrensemble() etc online documentation. Currently it is mentioned nowhere. Thank you both for your help!

(Bernhard, I am not sure if I can 'accept' your answer as THE answer, as it is a comment in Cris' answer. If you could please write it below, so that I can flag your answer as the accepted one..)