DALEX is designed to work with various black-box models like tree ensembles, linear models, neural networks etc. Unfortunately R packages that create such models are very inconsistent. Different tools use different interfaces to train, validate and use models.
In this vignette we will show explanations for models created with h2o package.
To illustrate applications of DALEX to regression problems we will use an artificial dataset apartments available in the DALEX package. Our goal is to predict the price per square meter of an apartment based on selected features such as construction year, surface, floor, number of rooms, district. It should be noted that four of these variables are continuous while the fifth one is a categorical one. Prices are given in Euro.
## m2.price construction.year surface floor no.rooms district
## 1 5897 1953 25 3 1 Srodmiescie
## 2 1818 1992 143 9 5 Bielany
## 3 3643 1937 56 1 2 Praga
## 4 3517 1995 93 7 3 Ochota
## 5 3013 1992 144 6 5 Mokotow
## 6 5795 1926 61 6 2 Srodmiescie
We create two regular H2O models: glm and gbm. To do this w need to first initialize h2o and we need to convert apartments to H2OFrame.
##
## H2O is not running yet, starting it now...
##
## Note: In case of errors look at the following log files:
## C:\Users\AnnA\AppData\Local\Temp\RtmpAFFy1j\file11c8b9b1e4/h2o_AnnA_started_from_r.out
## C:\Users\AnnA\AppData\Local\Temp\RtmpAFFy1j\file11c8576b2e7e/h2o_AnnA_started_from_r.err
##
##
## Starting H2O JVM and connecting: . Connection successful!
##
## R is connected to the H2O cluster:
## H2O cluster uptime: 6 seconds 350 milliseconds
## H2O cluster timezone: Europe/Belgrade
## H2O data parsing timezone: UTC
## H2O cluster version: 3.32.0.1
## H2O cluster version age: 1 month and 21 days
## H2O cluster name: H2O_started_from_R_AnnA_osn291
## H2O cluster total nodes: 1
## H2O cluster total memory: 3.48 GB
## H2O cluster total cores: 4
## H2O cluster allowed cores: 4
## H2O cluster healthy: TRUE
## H2O Connection ip: localhost
## H2O Connection port: 54321
## H2O Connection proxy: NA
## H2O Internal Security: FALSE
## H2O API Extensions: Amazon S3, Algos, AutoML, Core V3, TargetEncoder, Core V4
## R Version: R version 3.6.3 (2020-02-29)
h2o.no_progress()
apartments_hf <- as.h2o(apartments)
model_h2o_glm <- h2o.glm(y = "m2.price", training_frame = apartments_hf)
model_h2o_gbm <- h2o.gbm(y = "m2.price", training_frame = apartments_hf)
We also create third model by an H2O AutoML.
The first step of using the DALEX package is to wrap-up the black-box model with meta-data that unifies model interfacing. For model built in h2o
we can use DALEXtra, the extension of DALEX package.
To create an explainer we use explain_h2o()
function. Validation dataset for the models is apartmentsTest
data from the DALEX
package. For the models created by h2o package we have to provide custom predict function which takes two arguments: model
and newdata
and returns a numeric vector with predictions.
explainer_h2o_glm <- DALEXtra::explain_h2o(model = model_h2o_glm,
data = apartmentsTest[,2:6],
y = apartmentsTest$m2.price,
label = "h2o glm",
colorize = FALSE)
## Preparation of a new explainer is initiated
## -> model label : h2o glm
## -> data : 9000 rows 5 cols
## -> target variable : 9000 values
## -> predict function : yhat.H2ORegressionModel will be used ( default )
## -> predicted values : numerical, min = 2437.868 , mean = 3500.502 , max = 4694.49
## -> model_info : package h2o , ver. 3.32.0.1 , task regression ( default )
## -> residual function : difference between y and yhat ( default )
## -> residuals : numerical, min = -923.4434 , mean = 11.02155 , max = 2119.145
## A new explainer has been created!
explainer_h2o_gbm <- DALEXtra::explain_h2o(model = model_h2o_gbm,
data = apartmentsTest[,2:6],
y = apartmentsTest$m2.price,
label = "h2o gbm",
colorize = FALSE)
## Preparation of a new explainer is initiated
## -> model label : h2o gbm
## -> data : 9000 rows 5 cols
## -> target variable : 9000 values
## -> predict function : yhat.H2ORegressionModel will be used ( default )
## -> predicted values : numerical, min = 1702.853 , mean = 3505.4 , max = 6446.298
## -> model_info : package h2o , ver. 3.32.0.1 , task regression ( default )
## -> residual function : difference between y and yhat ( default )
## -> residuals : numerical, min = -587.5726 , mean = 6.123194 , max = 645.5172
## A new explainer has been created!
explainer_h2o_automl <- DALEXtra::explain_h2o(model = model_h2o_automl,
data = apartmentsTest[,2:6],
y = apartmentsTest$m2.price,
label = "h2o automl",
colorize = FALSE)
## Preparation of a new explainer is initiated
## -> model label : h2o automl
## -> data : 9000 rows 5 cols
## -> target variable : 9000 values
## -> predict function : yhat.H2ORegressionModel will be used ( default )
## -> predicted values : numerical, min = 1583.434 , mean = 3515.396 , max = 6570.365
## -> model_info : package h2o , ver. 3.32.0.1 , task regression ( default )
## -> residual function : difference between y and yhat ( default )
## -> residuals : numerical, min = -517.7739 , mean = -3.872876 , max = 571.5371
## A new explainer has been created!
Function model_performance()
calculates predictions and residuals for validation dataset.
mp_h2o_glm <- model_performance(explainer_h2o_glm)
mp_h2o_gbm <- model_performance(explainer_h2o_gbm)
mp_h2o_automl <- model_performance(explainer_h2o_automl)
Generic function print()
returns quantiles for residuals.
## Measures for: regression
## mse : 407698.6
## rmse : 638.5128
## r2 : 0.4971679
## mad : 481.9948
##
## Residuals:
## 0% 10% 20% 30% 40% 50% 60%
## -923.44339 -677.18027 -578.66955 -486.98678 -292.75927 -23.36066 85.25062
## 70% 80% 90% 100%
## 194.81788 426.24329 977.34524 2119.14517
Generic function plot()
shows reversed empirical cumulative distribution function for absolute values from residuals. Plots can be generated for one or more models.
We are also able to use the plot()
function to get an alternative comparison of residuals. Setting the geom = "boxplot"
parameter we can compare the distribution of residuals for selected models.
Using he DALEX package we are able to better understand which variables are important.
Model agnostic variable importance is calculated by means of permutations. We simply substract the loss function calculated for validation dataset with permuted values for a single variable from the loss function calculated for validation dataset.
This method is implemented in the model_parts()
function.
vi_h2o_glm <- model_parts(explainer_h2o_glm)
vi_h2o_gbm <- model_parts(explainer_h2o_gbm)
vi_h2o_automl <- model_parts(explainer_h2o_automl)
We can compare all models using the generic plot()
function.
Length of the interval coresponds to a variable importance. Longer interval means larger loss, so the variable is more important.
For better comparison of the models we can hook the variabe importance at 0 using the type="difference"
.
Explainers presented in this section are designed to better understand the relation between a variable and model output.
For more details of methods described in this section see Partial-dependence Profiles and Local-dependence and Accumulated-local Profiles
Partial Dependence Plots (PDP) are one of the most popular methods for exploration of the relation between a continuous variable and the model outcome.
Function model_profile()
with the parameter type = "partial"
calculates PDP response.
pdp_h2o_glm <- model_profile(explainer_h2o_glm, variable = "construction.year", type = "partial")
pdp_h2o_gbm <- model_profile(explainer_h2o_gbm, variable = "construction.year", type = "partial")
pdp_h2o_automl <- model_profile(explainer_h2o_automl, variable = "construction.year", type = "partial")
plot(pdp_h2o_glm, pdp_h2o_gbm, pdp_h2o_automl)
Acumulated Local Effects (ALE) plot is the extension of PDP, that is more suited for highly correlated variables.
Function model_profile()
with the parameter type = "accumulated"
calculate the ALE curve for the variable construction.year
.
ale_h2o_glm <- model_profile(explainer_h2o_glm, variable = "construction.year", type = "accumulated")
ale_h2o_gbm <- model_profile(explainer_h2o_gbm, variable = "construction.year", type = "accumulated")
ale_h2o_automl <- model_profile(explainer_h2o_automl, variable = "construction.year", type = "accumulated")
plot(ale_h2o_glm, ale_h2o_gbm, ale_h2o_automl)
Model prediction is visualized with Break Down Plots, which show the contribution of every variable present in the model. Function predict_parts()
with type = "break_down"
generates variable attributions for selected prediction. The generic plot()
function shows these attributions.
new_apartment <- apartmentsTest[1,]
pb_h2o_glm <- predict_parts(explainer_h2o_glm, new_observation = new_apartment, type = "break_down")
pb_h2o_gbm <- predict_parts(explainer_h2o_gbm, new_observation = new_apartment, type = "break_down")
pb_h2o_automl <- predict_parts(explainer_h2o_automl, new_observation = new_apartment, type = "break_down")
plot(pb_h2o_automl)
## R version 3.6.3 (2020-02-29)
## Platform: x86_64-w64-mingw32/x64 (64-bit)
## Running under: Windows 10 x64 (build 18363)
##
## Matrix products: default
##
## locale:
## [1] LC_COLLATE=Polish_Poland.1250 LC_CTYPE=Polish_Poland.1250
## [3] LC_MONETARY=Polish_Poland.1250 LC_NUMERIC=C
## [5] LC_TIME=Polish_Poland.1250
##
## attached base packages:
## [1] stats graphics grDevices utils datasets methods base
##
## other attached packages:
## [1] h2o_3.32.0.1 DALEX_2.0.1
##
## loaded via a namespace (and not attached):
## [1] Rcpp_1.0.4 pillar_1.4.3 compiler_3.6.3 ingredients_2.0
## [5] bitops_1.0-6 tools_3.6.3 bit_4.0.4 digest_0.6.25
## [9] jsonlite_1.6.1 evaluate_0.14 lifecycle_0.2.0 tibble_2.1.3
## [13] gtable_0.3.0 pkgconfig_2.0.3 rlang_0.4.6 yaml_2.2.1
## [17] xfun_0.12 DALEXtra_2.0 stringr_1.4.0 dplyr_1.0.0
## [21] knitr_1.28 generics_0.0.2 vctrs_0.3.1 bit64_4.0.5
## [25] grid_3.6.3 tidyselect_1.1.0 reticulate_1.14 glue_1.3.2
## [29] data.table_1.12.8 R6_2.4.1 iBreakDown_1.3.1 rmarkdown_2.1
## [33] farver_2.0.3 ggplot2_3.3.0 purrr_0.3.3 magrittr_1.5
## [37] scales_1.1.0 htmltools_0.4.0 colorspace_1.4-1 labeling_0.3
## [41] stringi_1.4.6 RCurl_1.98-1.2 munsell_0.5.0 crayon_1.3.4