shapleyR is an R package that provides some functionality to use mlr tasks and models to generate shapley values. And thus analyze the effects of the features on the outcome of a model.
shapleyR already supports the regression, classification, clustering and multilabel tasks from mlr. We plan to add the missing tasks from that package.
The package can be installed directly from github with devtools (see following section). Beside that we also plan to upload this package to CRAN as soon as it gets production ready.
install.packages("devtools")
devtools::install_github('redichh/shapleyR')
library(shapleyR)
As a quickstart we will calculate the shapley values for a regression task. For that we take a look at the Boston Housing dataset. This is alredy included in the mlr-package and can be called with bh.task from the R terminal.
The Dataset looks as following:
> head(getTaskData(bh.task))
crim zn indus chas nox rm age dis rad tax ptratio b lstat medv
0.00632 18 2.31 0 0.538 6.575 65.2 4.0900 1 296 15.3 396.90 4.98 24.0
0.02731 0 7.07 0 0.469 6.421 78.9 4.9671 2 242 17.8 396.90 9.14 21.6
0.02729 0 7.07 0 0.469 7.185 61.1 4.9671 2 242 17.8 392.83 4.03 34.7
0.03237 0 2.18 0 0.458 6.998 45.8 6.0622 3 222 18.7 394.63 2.94 33.4
0.06905 0 2.18 0 0.458 7.147 54.2 6.0622 3 222 18.7 396.90 5.33 36.2
0.02985 0 2.18 0 0.458 6.430 58.7 6.0622 3 222 18.7 394.12 5.21 28.7
Wheras the first 13 features are the describing variables and the last one "medv" is the dependant variable. The prediction for the medv feature is:
> prediction = head(getPredictionResponse(predict(train("regr.lm", bh.task), newdata = getTaskData(bh.task))))
[1] 30.00384 25.02556 30.56760 28.60704 27.94352 25.25628
And the mean for medv over all observation available in the dataset is:
> data.mean = mean(getTaskData(bh.task)[,getTaskTargetNames(bh.task)])
[1] 22.53281
Now we can take a look at the shapley-function itself. Running following code shows the influence for every feature according to a specific observation:
> shap.values = getShapleyValues(shapley(1:6, task = bh.task, model = train("regr.lm", bh.task)))
_Id _Class crim zn indus chas nox rm age dis rad tax ptratio b lstat
1 1 NA 0.275 0.256 -0.185 -0.358 1.070 1.632 -0.002 0.065 -2.510 1.229 2.690 0.337 4.468
2 2 NA 0.503 -0.385 -0.078 -0.358 2.026 -0.014 0.005 -2.157 -1.969 2.419 0.407 0.455 2.233
3 3 NA 0.389 -0.244 -0.090 -0.269 1.542 3.346 -0.008 -1.740 -2.724 2.139 0.213 0.227 4.538
4 4 NA 0.363 -0.390 -0.152 -0.090 1.673 2.048 -0.014 -3.148 -1.938 1.746 -0.254 0.318 4.894
5 5 NA 0.427 -0.666 -0.167 -0.448 1.840 3.286 -0.012 -3.435 -3.060 2.852 0.079 0.450 4.004
6 6 NA 0.454 -0.430 -0.165 0.000 1.957 1.211 -0.010 -3.336 -1.367 2.455 -0.143 0.666 3.907
Taking the sum of all explaining features for every row results in the following:
> approximation = rowSums(shap.values[,getTaskFeatureNames(bh.task)])
[1] 8.967 3.087 7.319 5.056 5.150 5.199
And this is the approximated difference between the previously calculated prediction and data.mean. Assuming that the sum of all shapley values for one observation equals the difference between the prediction and the data mean the following calculation should be close to zero:
> prediction - data.mean - approximation
[1] -1.4959629 -0.5942439 0.7157904 1.0182302 0.2607179 -2.4755219
We see that this is not the case. But increasing the amount of iterations should lead to better results:
> shap.values.2 = getShapleyValues(shapley(1:6, task = bh.task, model = train("regr.lm", bh.task), iterations = 200))
> approximation.2 = rowSums(shap.values.2[,getTaskFeatureNames(bh.task)])
> prediction - data.mean - approximation.2
[1] -0.1209629 0.7367561 -0.3352096 -0.1327698 -0.1692821 -0.1395219
tbd
In our Vignette can be found further information about this package. There is also shown the usage of plots for the shapley values.