Exploring models with DALEX
An issue with machine learning algorithms is that it is not easy to
understand the role of different variables in giving the final
prediction. A number of packages have been created to explore and
explain the behaviour of ML algorithms, such as those used in
tidysdm. In the tidysdm
overview article, we illustrated how to use recipes to
create profiles.
Here we demonstrate how to use DALEX, an excellent
package that has methods to deal with tidymodels.
tidysdm contains additional functions that allow use to use
the DALEX functions directly on tidysdm ensembles.
We will use a simple ensemble that we built in the
library(tidysdm)
lacerta_ensemble
#> A simple_ensemble of models
#>
#> Members:
#> • default_glm
#> • default_rf
#> • default_gbm
#> • default_maxent
#>
#> Available metrics:
#> • boyce_cont
#> • roc_auc
#> • tss_max
#>
#> Metric used to tune workflows:
#> • boyce_cont
The first step in DALEX is to create an explainer object, which can
then be queried by different functions in the package, to turn the
explainer into an explanation (following the DALEX lingo). As a first
step, we use the custom function explain_tidysdm to
generate our explainer:
explainer_lacerta_ens <- explain_tidysdm(lacerta_ensemble)
#> Preparation of a new explainer is initiated
#> -> model label : data.frame ( default )
#> -> data : 444 rows 4 cols
#> -> data : tibble converted into a data.frame
#> -> target variable : 444 values
#> -> predict function : predict_function
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package tidysdm , ver. 0.9.5 , task classification ( default )
#> -> model_info : type set to classification
#> -> predicted values : numerical, min = 0.02368574 , mean = 0.2952422 , max = 0.7489003
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.6793936 , mean = -0.04524217 , max = 0.7953426
#> A new explainer has been created!
Now that we have our explainer, we can explore variable importance
for the ensemble:
library(DALEX)
#> Welcome to DALEX (version: 2.4.3).
#> Find examples and detailed introduction at: http://ema.drwhy.ai/
#>
#> Attaching package: 'DALEX'
#> The following object is masked from 'package:dplyr':
#>
#> explain
vip_ensemble <- model_parts(explainer = explainer_lacerta_ens)
plot(vip_ensemble)
Or generate partial dependency plots for a given variable
(e.g. bio05):
pdp_bio05 <- model_profile(explainer_lacerta_ens, N = 500, variables = "bio05")
plot(pdp_bio05)
There are many other functions in DALEX that can be applied to the
explainer to further explore the behaviour of the model; see several
tutorial on https://modeloriented.github.io/DALEX/
It is also possible to explore the individual models that make up the
ensemble:
explainer_list <- explain_tidysdm(tidysdm::lacerta_ensemble, by_workflow = TRUE)
#> Preparation of a new explainer is initiated
#> -> model label : default_glm
#> -> data : 444 rows 4 cols
#> -> data : tibble converted into a data.frame
#> -> target variable : 444 values
#> -> predict function : yhat.workflow will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package tidymodels , ver. 1.2.0 , task classification ( default )
#> -> model_info : type set to classification
#> -> predicted values : numerical, min = 0.005923341 , mean = 0.75 , max = 0.9993179
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.9513805 , mean = 3.739236e-15 , max = 0.9628654
#> A new explainer has been created!
#> Preparation of a new explainer is initiated
#> -> model label : default_rf
#> -> data : 444 rows 4 cols
#> -> data : tibble converted into a data.frame
#> -> target variable : 444 values
#> -> predict function : yhat.workflow will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package tidymodels , ver. 1.2.0 , task classification ( default )
#> -> model_info : type set to classification
#> -> predicted values : numerical, min = 0 , mean = 0.7514254 , max = 1
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.5907905 , mean = -0.001425431 , max = 0.523096
#> A new explainer has been created!
#> Preparation of a new explainer is initiated
#> -> model label : default_gbm
#> -> data : 444 rows 4 cols
#> -> data : tibble converted into a data.frame
#> -> target variable : 444 values
#> -> predict function : yhat.workflow will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package tidymodels , ver. 1.2.0 , task classification ( default )
#> -> model_info : type set to classification
#> -> predicted values : numerical, min = 0.255644 , mean = 0.6272817 , max = 0.7515493
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.7198042 , mean = 0.1227183 , max = 0.582286
#> A new explainer has been created!
#> Preparation of a new explainer is initiated
#> -> model label : default_maxent
#> -> data : 444 rows 4 cols
#> -> data : tibble converted into a data.frame
#> -> target variable : 444 values
#> -> predict function : yhat.workflow will be used ( default )
#> -> predicted values : No value for predict function target column. ( default )
#> -> model_info : package tidymodels , ver. 1.2.0 , task classification ( default )
#> -> model_info : type set to classification
#> -> predicted values : numerical, min = 0.09570046 , mean = 0.7996343 , max = 0.9997615
#> -> residual function : difference between y and yhat ( default )
#> -> residuals : numerical, min = -0.8792301 , mean = -0.04963429 , max = 0.6252604
#> A new explainer has been created!
The resulting list can be then used to build lists of explanations,
which can then be plotted.
profile_list <- lapply(explainer_list, model_profile,
N = 500,
variables = "bio05"
)
plot(profile_list)
The initial split
The standard approach in tidymodels is to make an
initial split of the data into a test and a training set. We will use
retain 20% of the data (1/5) for the testing set, and use the rest for
training.
We start by loading a set of presences and absences and their
associated climate, analogous to the one that we generated in the tidysdm
overview article:
library(tidysdm)
library(sf)
lacerta_thin <- readRDS(system.file("extdata/lacerta_climate_sf.RDS",
package = "tidysdm"
))
We then use spatial_initial_split to do the split, using
a spatial_block_cv scheme to partition the data:
set.seed(1005)
lacerta_initial <- spatial_initial_split(lacerta_thin,
prop = 1 / 5, spatial_block_cv
)
autoplot(lacerta_initial)
And check the balance of presences vs pseudoabsences:
check_splits_balance(lacerta_initial, class)
#> # A tibble: 1 × 4
#> presence_test pseudoabs_test presence_train pseudoabs_train
#> <int> <int> <int> <int>
#> 1 88 267 25 72
We can now extract the training set from our
lacerta_initial split, and sample folds to set up cross
validation (note that we set the cellsize and
offset based on the full dataset,
lacerta_thin; this allows us to use the same grid we used
for the initial_split).
set.seed(1005)
lacerta_training <- training(lacerta_initial)
lacerta_cv <- spatial_block_cv(lacerta_training,
v = 5,
cellsize = grid_cellsize(lacerta_thin),
offset = grid_offset(lacerta_thin)
)
autoplot(lacerta_cv)
And check the balance in the dataset:
check_splits_balance(lacerta_cv, class)
#> # A tibble: 5 × 4
#> presence_assessment pseudoabs_assessment presence_analysis pseudoabs_analysis
#> <int> <int> <int> <int>
#> 1 74 197 14 70
#> 2 59 225 29 42
#> 3 73 220 15 47
#> 4 76 209 12 58
#> 5 70 218 18 49
Different recipes for certain models
Only certain type of models (e.g. glm, svm) struggle with correlated
variables; other algorithms, such as random forests, can handle
correlated variables. So, we will create two recipes, one with all
variables, and one only with the variables that are uncorrelated:
lacerta_rec_all <- recipe(lacerta_thin, formula = class ~ .)
lacerta_rec_uncor <- lacerta_rec_all %>%
step_rm(all_of(c(
"bio01", "bio02", "bio03", "bio04", "bio07", "bio08",
"bio09", "bio10", "bio11", "bio12", "bio14", "bio16",
"bio17", "bio18", "bio19", "altitude"
)))
lacerta_rec_uncor
#>
#> ── Recipe ──────────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> outcome: 1
#> predictor: 20
#> coords: 2
#>
#> ── Operations
#> • Variables removed: all_of(c("bio01", "bio02", "bio03", "bio04", "bio07",
#> "bio08", "bio09", "bio10", "bio11", "bio12", "bio14", "bio16", "bio17",
#> "bio18", "bio19", "altitude"))
And now use these two recipes in a workflowset (we will
keep it small for computational time), selecting the appropriate recipe
for each model. We will include a model (polynomial support vector
machines, or SVM) which does not have a wrapper in tidysdm
for creating a model specification. However, we can use a standard model
spec from yardstick:
lacerta_models <-
# create the workflow_set
workflow_set(
preproc = list(
uncor = lacerta_rec_uncor, # recipe for the glm
all = lacerta_rec_all, # recipe for the random forest
all = lacerta_rec_uncor # recipe for svm
),
models = list(
# the standard glm specs
glm = sdm_spec_glm(),
# rf specs with tuning
rf = sdm_spec_rf(),
# svm specs with tuning
svm = parsnip::svm_poly(
cost = tune(),
degree = tune()
) %>%
parsnip::set_engine("kernlab") %>%
parsnip::set_mode("classification")
),
# make all combinations of preproc and models,
cross = FALSE
) %>%
# tweak controls to store information needed later to create the ensemble
# note that we use the bayes version as we will use a Bayes search (see later)
option_add(control = stacks::control_stack_bayes())
We can now use the block CV folds to tune and assess the models. Note
that there are multiple tuning approaches, besides the standard grid
method. Here we will use tune_bayes from the
tune package (see its help page to see how a Gaussian
Process model is used to choose parameter combinations).
This tuning method (as opposed to use a standard grid) does not allow
for hyper-parameters with unknown limits, but mtry for
random forest is undefined as its upper range depends on the number of
variables in the dataset. So, before tuning, we need to finalise
mtry by informing the set dials with the actual data:
rf_param <- lacerta_models %>%
# extract the rf workflow
extract_workflow("all_rf") %>%
# extract its parameters dials (used to tune)
extract_parameter_set_dials() %>%
# give it the predictors to finalize mtry
finalize(x = st_drop_geometry(lacerta_thin) %>% select(-class))
# now update the workflowset with the new parameter info
lacerta_models <- lacerta_models %>%
option_add(param_info = rf_param, id = "all_rf")
And now we can tune the models:
set.seed(1234567)
lacerta_models <-
lacerta_models %>%
workflow_map("tune_bayes",
resamples = lacerta_cv, initial = 8,
metrics = sdm_metric_set(), verbose = TRUE
)
#> i No tuning parameters. `fit_resamples()` will be attempted
#> i 1 of 3 resampling: uncor_glm
#> ✔ 1 of 3 resampling: uncor_glm (406ms)
#> i 2 of 3 tuning: all_rf
#> ! No improvement for 10 iterations; returning current results.
#> ✔ 2 of 3 tuning: all_rf (22s)
#> i 3 of 3 tuning: all_svm
#> ✔ 3 of 3 tuning: all_svm (29.3s)
We can have a look at the performance of our models with:
Stack ensembles
Instead of building a simple ensemble with the best version of each
model type, we can build a stack ensemble, as implemented in the package
stacks. Stacking uses a meta-learning algorithm to learn
how to best combine multiple models, including multiple versions of the
same algorithm with different hyper-parameters.
library(stacks)
set.seed(1005)
lacerta_stack <-
# initialize the stack
stacks() %>%
# add candidate members
add_candidates(lacerta_models) %>%
# determine how to combine their predictions
blend_predictions() %>%
# fit the candidates with non-zero weights (i.e.non-zero stacking coefficients)
fit_members()
autoplot(lacerta_stack, type = "weights")
We can see that three versions of the SVM and one of the random
forests were selected; the stacking coefficients give an indication of
the weight each model carries within the ensemble. We can now use the
ensemble to make predictions about the testing data:
lacerta_testing <- testing(lacerta_initial)
lacerta_test_pred <-
lacerta_testing %>%
bind_cols(predict(lacerta_stack, ., type = "prob"))
And look at the goodness of fit using some commonly used sdm metrics.
Note that sdm_metric_set is first invoked to generate a
function (with empty ()) that is then used on the data.
sdm_metric_set()(data = lacerta_test_pred, truth = class, .pred_presence)
#> # A tibble: 3 × 3
#> .metric .estimator .estimate
#> <chr> <chr> <dbl>
#> 1 boyce_cont binary 0.853
#> 2 roc_auc binary 0.986
#> 3 tss_max binary 0.92
We can now make predictions with this stacked ensemble. We start by
extracting the climate for the variables of interest
download_dataset("WorldClim_2.1_10m")
climate_vars <- lacerta_rec_all$var_info %>%
filter(role == "predictor") %>%
pull(variable)
climate_present <- pastclim::region_slice(
time_ce = 1985,
bio_variables = climate_vars,
data = "WorldClim_2.1_10m",
crop = iberia_poly
)
prediction_present <- predict_raster(lacerta_stack, climate_present,
type = "prob"
)
library(tidyterra)
ggplot() +
geom_spatraster(data = prediction_present, aes(fill = .pred_presence)) +
scale_fill_terrain_c() +
# plot presences used in the model
geom_sf(data = lacerta_thin %>% filter(class == "presence"))
