What is the weights argument for in the R gbm function? Does it implement cost-sensitive stochastic gradient boosting?
You may have already read this, but the documentation says that the weights parameter is defined in this way:
an optional vector of weights to be used in the fitting process. Must
be positive but do not need to be normalized. If keep.data=FALSE in
the initial call to gbm then it is the user’s responsibility to
resupply the weights to gbm.more.
Thus my interpretation would be that they are standard observation weights as in any statistical model.
Is it cost-sensitive? Good question. I first noticed that one of the main citations for the package is:
B. Kriegler (2007). Cost-Sensitive Stochastic Gradient Boosting Within a Quantitative Regression Framework.
so I figured it does imply cost-sensitivity, but there's not an explicit use of that term in the vignette, so if it was not apparent.
I did a little bit of a deeper dive though and found some more resources. You can find the equations describing the weights towards the end of this article which describes the package.
I also found this question being asked way back in 2009 in a mailing list, and while there was no response, I finally found a a scholarly article discussing the use of gbm and other R packages for cost-sensitive gradient boosting.
The conclusion is that gbm's quantile loss function is differentiable and can be used in cost-sensitive applications wherein over/under-estimation have different error costs, however other quantitative loss functions (aside from quantile) may be necessary/appropriate in some applications of cost-sensitive gradient boosting.
That paper is centered around gbm but also discusses other packages and if your focus is on cost-sensitive gradient boosting then you may want to look at the others they mention in the paper as well.
Related
I try to optimize the averaged prediction of two logistic regressions in a classification task using a superlearner.
My measure of interest is classif.auc
The mlr3 help file tells me (?mlr_learners_avg)
Predictions are averaged using weights (in order of appearance in the
data) which are optimized using nonlinear optimization from the
package "nloptr" for a measure provided in measure (defaults to
classif.acc for LearnerClassifAvg and regr.mse for LearnerRegrAvg).
Learned weights can be obtained from $model. Using non-linear
optimization is implemented in the SuperLearner R package. For a more
detailed analysis the reader is referred to LeDell (2015).
I have two questions regarding this information:
When I look at the source code I think LearnerClassifAvg$new() defaults to "classif.ce", is that true?
I think I could set it to classif.auc with param_set$values <- list(measure="classif.auc",optimizer="nloptr",log_level="warn")
The help file refers to the SuperLearner package and LeDell 2015. As I understand it correctly, the proposed "AUC-Maximizing Ensembles through Metalearning" solution from the paper above is, however, not impelemented in mlr3? Or do I miss something? Could this solution be applied in mlr3? In the mlr3 book I found a paragraph regarding calling an external optimization function, would that be possible for SuperLearner?
As far as I understand it, LeDell2015 proposes and evaluate a general strategy that optimizes AUC as a black-box function by learning optimal weights. They do not really propose a best strategy or any concrete defaults so I looked into the defaults of the SuperLearner package's AUC optimization strategy.
Assuming I understood the paper correctly:
The LearnerClassifAvg basically implements what is proposed in LeDell2015 namely, it optimizes the weights for any metric using non-linear optimization. LeDell2015 focus on the special case of optimizing AUC. As you rightly pointed out, by setting the measure to "classif.auc" you get a meta-learner that optimizes AUC. The default with respect to which optimization routine is used deviates between mlr3pipelines and the SuperLearner package, where we use NLOPT_LN_COBYLA and SuperLearner ... uses the Nelder-Mead method via the optim function to minimize rank loss (from the documentation).
So in order to get exactly the same behaviour, you would need to implement a Nelder-Mead bbotk::Optimizer similar to here that simply wraps stats::optim with method Nelder-Mead and carefully compare settings and stopping criteria. I am fairly confident that NLOPT_LN_COBYLA delivers somewhat comparable results, LeDell2015 has a comparison of the different optimizers for further reference.
Thanks for spotting the error in the documentation. I agree, that the description is a little unclear and I will try to improve this!
Thank you for seeing this post.
Various regression models are being applied to the curve estimating (actual measured ventilation rate).
Comparison was made using the GLM and GAM models including polynomial regression. I use R.
Are there any other types of regression models that can simulate that curve well?
like...bayesian? (In fact, I didn't even understand if it could be applied here)
Sincerely.
loads of non linear methods exist, and many would give similar results, but this is a statistics question. it would fit better on cross validated. also, you have to tell: do you want to do interpolation, even extrapolation? what is your analysis for?
bayesian methods are used when you have knowledge of the phenomenon prior to data, or in some cases when you want to apply regularization or graphical models to data generation processes, I think you should better leave out bayesian statistics here.
edit:
to be short: if you want to obtain a readable formulation of the curve, and you don't have any specific mathematical model in mind, go for a polynomial fit. Other popular choices, which are better for only plotting the curve, instead of reporting it in a mathematical expression, are smoothing splines and LOESS. for further details, maybe ask a new question on stats.stackexchange.com, after studing better the alternatives.
"Gradient boosting", like AdaBoost is built for any weak learner. In that sense a "gradient boosting machine" might be thought of as a generic wrapper, and not necessarily a wrapper for a classification and regression tree (CART) model. When I say "CART", and am referring to the regression, I am looking for something outside a constant (mean) or a linear or generalized linear fit.
Are there any R libraries that allow me to wrap functions into sequential gradient boosting? I'm not looking at "better vs. worse", I'm looking for "minimally capable".
I know that there are many that allow me to make a long sequence of gradient boosted CART models and I am not looking for those.
Libraries (afaict) that can't do what I want:
gbm - CART
xgboost - CART
GAMboost - B-splines
Notes:
I'm still looking and will report results that I find.
If my question has the wrong tags, and you think of one that is more apropos, please feel free to share.
This is a follow-up to a previous question I asked a while back that was recently answered.
I have built several gbm models with dismo::gbm.step, which relies on the gbm fitting functions found in R package gbm, as well as cross validation tools from R package splines.
As part of my analysis, I would like to use some of the graphical tools available in R (e. g. perspective plots) to visualize pairwise interactions in the data. Both the gbm and the dismo packages have functions for detecting and modelling interactions in the data.
The implementation in dismo is explained in Elith et. al (2008) and returns a statistic which indicates departures of the model predictions from a linear combination of the predictors, while holding all other predictors at their means.
The implementation in gbm uses Friedman`s H statistic (Friedman & Popescue, 2005), and returns a different metric, and also does NOT set the other variables at their means.
The interactions modelled and plotted with dismo::gbm.interactions are great and have been very informative. However, I would also like to use gbm::interact.gbm, partly for publication strength and also to compare the results from the two methods.
If I try to run gbm::interact.gbm in a gbm.object created with dismo, an error is returned…
"Error in is.factor(data[, x$var.names[j]]) :
argument "data" is missing, with no default"
I understand dismo::gmb.step adds extra data the authors thought would be useful to the gbm model.
I also understand that the answer to my question lies somewherein the source code.
My questions is...
Is it possible to modify a gbm object created in dismo to be used in gbm::gbm.interact? If so, would this be accomplished by...
a. Modifying the gbm object created in dismo::gbm.step?
b. Modifying the source code for gbm::interact.gbm?
c. Doing something else?
I will be going through the source code trying to solve this myself, if I come up with a solution before anyone answers I will answer my own question.
The gbm::interact.gbm function requires data as an argument interact.gbm <- function(x, data, i.var = 1, n.trees = x$n.trees).
The dismo gbm.object is essentially the same as the gbm gbm.object, but with extra information attached so I don't imagine changing the gbm.object would help.
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
This question does not appear to be about programming within the scope defined in the help center.
Closed 1 year ago.
Improve this question
I am doing regression task - do I need to normalize (or scale) data for randomForest (R package)? And is it neccessary to scale also target values?
And if - I want to use scale function from caret package, but I did not find how to get data back (descale, denormalize). Do not you know about some other function (in any package) which is helpfull with normalization/denormalization?
Thanks,
Milan
No, scaling is not necessary for random forests.
The nature of RF is such that convergence and numerical precision issues, which can sometimes trip up the algorithms used in logistic and linear regression, as well as neural networks, aren't so important. Because of this, you don't need to transform variables to a common scale like you might with a NN.
You're don't get any analogue of a regression coefficient, which measures the relationship between each predictor variable and the response. Because of this, you also don't need to consider how to interpret such coefficients which is something that is affected by variable measurement scales.
Scaling is done to Normalize data so that priority is not given to a particular feature.
Role of Scaling is mostly important in algorithms that are distance based and require Euclidean Distance.
Random Forest is a tree-based model and hence does not require feature scaling.
This algorithm requires partitioning, even if you apply Normalization then also> the result would be the same.
I do not see any suggestions in either the help page or the Vignette that suggests scaling is necessary for a regression variable in randomForest. This example at Stats Exchange does not use scaling either.
Copy of my comment: The scale function does not belong to pkg:caret. It is part of the "base" R package. There is an unscale function in packages grt and DMwR that will reverse the transformation, or you could simply multiply by the scale attribute and then add the center attribute values.
Your conception of why "normalization" needs to be done may require critical examination. The test of non-normality is only needed after the regressions are done and may not be needed at all if there are no assumptions of normality in the goodness of fit methodology. So: Why are you asking? Searching in SO and Stats.Exchange might prove useful:
citation #1 ; citation #2 ; citation #3
The boxcox function is a commonly used tranformation when one does not have prior knowledge of twhat a distribution "should" be and when you really need to do a tranformation. There are many pitfalls in applying transformations, so the fact that you need to ask the question raises concerns that you may be in need of further consultations or self-study.
Guess, what will happen in the following example?
Imagine, you have 20 predictive features, 18 of them are in [0;10] range and the other 2 in [0;1,000,000] range (taken from a real-life example). Question1: what feature importances will Random Forest assign. Question2: what will happen to the feature importance after scaling the 2 large-range features?
Scaling is important. It is that Random Forest is less sensitive to the scaling then other algorithms and can work with "roughly"-scaled features.
If you are going to add interactions to dataset - that is, new variable being some function of other variables (usually simple multiplication), and you dont feel what that new variable stands for (cant interprete it), then you should calculate this variable using scaled variables.
Random Forest uses information gain / gini coefficient inherently which will not be affected by scaling unlike many other machine learning models which will (such as k-means clustering, PCA etc). However, it might 'arguably' fasten the convergence as hinted in other answers