Modeling Options

Modeling options

biomod2 is a wrapper calling for single models functions from external packages. Modeling options are automatically retrieved from these packages, allowing the use of all arguments taken into account by these functions.
Default parameter values are unmodified and often non-adapted to species distribution modeling in general, and to specific dataset in particular. Bigboss options provided by biomod2 team tend to correct at least the species distribution modeling aspect, while tuned options allow to try and find more appropriate parameterization for user data through caret package mainly. The user can also defines its own modeling options parameterization (user.defined).

Note that only binary data type and associated models are allowed currently, but the package structure has been changed to enable the addition of new data types in near future, such as absolute or relative abundances.

In the dataset ModelsTable, all the different algorithms are listed with their packages and functions :

           model   type      package         func       train
1            ANN binary         nnet         nnet      avNNet
2            CTA binary        rpart        rpart       rpart
3            FDA binary          mda          fda         fda
4            GAM binary          gam          gam    gamLoess
5            GAM binary         mgcv          bam         bam
6            GAM binary         mgcv          gam         gam
7            GBM binary          gbm          gbm         gbm
8            GLM binary        stats          glm         glm
9           MARS binary        earth        earth       earth
10        MAXENT binary       MAXENT       MAXENT ENMevaluate
11        MAXNET binary       maxnet       maxnet      maxnet
12            RF binary randomForest randomForest          rf
13           SRE binary      biomod2       bm_SRE      bm_SRE
14       XGBOOST binary      xgboost      xgboost     xgbTree

All the examples are made with the data of the package.
For the beginning of the code, see the main functions vignette.


Default options

biomod2 has a set of default options, matching most of the time the algorithms’ default values, but with some minor modifications to allow the BIOMOD_Modeling function to run smoothly.

Please be aware that this strategy can often lead to bad models or even some errors.

myBiomodModelOut <- BIOMOD_Modeling(bm.format = myBiomodData,
                                    modeling.id = 'Example',
                                    models = c('RF', 'GLM'),
                                    CV.strategy = 'random',
                                    CV.nb.rep = 2,
                                    CV.perc = 0.8,
                                    OPT.strategy = 'default',
                                    metric.eval = c('TSS','ROC'),
                                    var.import = 2,
                                    seed.val = 42)

You can retrieve the models options with get_options

get_options(myBiomodModelOut)

Bigboss options

The bigboss set of parameters is available in the dataset OptionsBigboss. This set should give better results than the default set and will be continued to be optimized by the biomod2 Team.

Keep in mind that this is something general and dependent of your case, the results can be not better than the default set.

myBiomodModelOut <- BIOMOD_Modeling(bm.format = myBiomodData,
                                    modeling.id = 'Example',
                                    models = c('RF', 'GLM'),
                                    CV.strategy = 'random',
                                    CV.nb.rep = 2,
                                    CV.perc = 0.8,
                                    OPT.strategy = 'bigboss',
                                    metric.eval = c('TSS','ROC'),
                                    var.import = 2,
                                    seed.val = 42)

Tuned options

With tuned options, some algorithms can be trained over your dataset, and optimized parameters are returned to be used within the BIOMOD_Modeling function. This tuning is mostly based upon the caret package which calls a specific function to tune each algorithm (see column train in ModelsTable). As exception, the ENMevaluate function of the ENMeval package is called for MAXENT and the biomod2 team wrote a special function for SRE.

Here is the list of the parameters that can be tuned :

algorithm parameters
ANN size, decay, bag
FDA degree, nprune
GAM select, method
GBM n.trees, interaction.depth, shrinkage, n.minobsinnode
MARS degree, nprune
RF mtry
SRE quant
XGBOOST nrounds, max_depth, eta, gamma, colsampl_bytree, min_child_weight, subsample

For almost every algorithm (except MAXENT, MAXNET and SRE), you can choose to optimize the formula by setting do.formula = TRUE. The optimized formula will be chosen between the different type (simple, quadratic, polynomial, s_smoother) and for different interaction level.
In the same way, a variable selection can be run for GLM and GAM if do.stepAIC = TRUE (respectively, MASS::stepAIC and gam::step.Gam).

More information about the training can be found in the documentation of the bm_Tuning function.

myBiomodModelOut <- BIOMOD_Modeling(bm.format = myBiomodData,
                                    modeling.id = 'Example',
                                    models = c('RF','SRE'),
                                    CV.strategy = 'random',
                                    CV.nb.rep = 2,
                                    CV.perc = 0.8,
                                    OPT.strategy = 'tuned',
                                    metric.eval = c('TSS','ROC'),
                                    var.import = 2,
                                    seed.val = 42)

print(get_options(myBiomodModelOut), dataset = '_allData_RUN1')                                    

User defined

The user.defined option allows you to adjust yourself the parameters of all the algorithms.

Note that you can find information about the parameters of MAXENT within the documentation of the bm_ModelingOptions function.


Example :

myCVtable <- bm_CrossValidation(bm.format = myBiomodData,
                                strategy = "random",
                                nb.rep = 2,
                                perc = 0.8)


myOpt  <- bm_ModelingOptions(data.type = 'binary',
                             models = c('RF','GLM','MARS'),
                             strategy = 'bigboss',
                             bm.format = myBiomodData, 
                             calib.lines = myCVtable)

print(myOpt)
tuned.rf <- bm_Tuning(model = 'RF',
                      tuning.fun = 'rf', ## see in ModelsTable
                      do.formula = TRUE,
                      bm.options = opt.d@options$RF.binary.randomForest.randomForest,
                      bm.format = myBiomodData, 
                      calib.lines = myCVtable)

form.GLM <- bm_MakeFormula(resp.name = myBiomodData@sp.name,
                           expl.var = head(myBiomodData@data.env.var),
                           type = 'simple',
                           interaction.level = 0)
                     
user.GLM <- list('_allData_RUN1' = list(formula = form.GLM),
                 '_allData_RUN2' = list(formula = form.GLM))

## Gather in one list
## Models names can be found in OptionsBigboss@models
user.val <- list( RF.binary.randomForest.randomForest = tuned.rf,
                  GLM.binary.stats.glm= user.GLM)

myOpt <- bm_ModelingOptions(data.type = 'binary',
                            models = c('RF','GLM','MARS'),
                            strategy = "user.defined",
                            user.val = user.val,
                            user.base = "bigboss",
                            bm.format = myBiomodData, 
                            calib.lines = myCVtable)
                            
print(myOpt)
print(myOpt, dataset = '_allData_RUN1')
print(myOpt, dataset = '_allData_RUN2')

myBiomodModelOut <- BIOMOD_Modeling(bm.format = myBiomodData,
                                modeling.id = 'Example',
                                models = c('RF','GLM','MARS'),
                                CV.strategy = 'user.defined',
                                CV.user.table = myCVtable,
                                OPT.user = myOpt,
                                metric.eval = c('TSS','ROC'),
                                var.import = 2)

You can find more examples in the Secondary functions vignette.