Why always the same prediction error when using LSBoost with fitrensemble?

Tobias Pahlberg

2017 4 23

0 个回答

3 次查看（30 天）

0 个投票

I was using a trained regression ensemble on some test data when I realized that the prediction error was the same every time I ran it. Usually, since I have not fixed the random number generator seed, I get slightly different results. I figured out that if I change the 'NumVariablesToSample' from the default: 'all', I get different results. But why? And is it a better practice to run with slightly less variables just to get a mean and SD of the prediction error?

Here is the code that I run (here with added NumVariablesToSample)

tempTree = templateTree('MaxNumSplits',2,'MinLeafSize',7, 'NumVariablesToSample',2262);
ens = fitrensemble(trainData,'FVPROCSTP','NumLearningCycles',364,...
        'Method', 'LSBoost', 'Learners',tempTree,'LearnRate',0.11779);

/Tobias

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Roisin Loughnane 2017-10-9

Your question did not get answered. Did you manage to solve your problem? As I am looking for answers to similar questions right now. I do not have access to fitrensemble, only fitensemble. With LSBoost can you model the data with a logit (probability) error distribution do you know?

Also, you have very specific learner rate, no. of cycle parameters.. How did you optimise these?

请先登录，再进行评论。

请先登录，再回答此问题。

Follow Question