Explaining nestedcv models with Shapley values

nestedcv provides two methods for understanding fitted models. The simplest of these is to plot variable importance. The newer method is to calculate Shapley values for each predictor.

Variable importance and variable stability

For regression model systems such as glmnet variable importance is represented as the coefficients of the model scaled by absolute value from largest to smallest. However, the outer folds of nested CV allow us to show the variance of model coefficients across each outer fold plus the final model, and hence see how stable the model is. We can also overlay how often predictors are selected in each model to give a sense of the stability of predictor selection.

In the example below using the Boston housing dataset, a glmnet regression model is fitted and the variable importance for predictors is shown based on the coefficients in the final model and tuned models from 10 outer folds. min_1se is set to 1, which is the equivalent of specifying s = "lambda.1se" with glmnet, to encourage a more sparse model.

library(nestedcv)
library(mlbench)  # Boston housing dataset

data(BostonHousing2)
dat <- BostonHousing2
y <- dat$cmedv
x <- subset(dat, select = -c(cmedv, medv, town, chas))

# Fit a glmnet model using nested CV
set.seed(1, "L'Ecuyer-CMRG")
fit <- nestcv.glmnet(y, x, family = "gaussian",
                     min_1se = 1, alphaSet = 1, cv.cores = 2)
vs <- var_stability(fit)
vs
##                 mean           sd         sem frequency sign direction final
## lon     48.393372145 13.359312172 4.027984176        11   -1  negative   yes
## rm      35.277524448 12.364884451 3.728152936        11    1  positive   yes
## ptratio  5.660722405  1.767597466 0.532950689        11   -1  negative   yes
## lstat    4.460941717  1.651898823 0.498066235        11   -1  negative   yes
## nox      3.632481722  5.345960957 1.611867876         4   -1  negative   yes
## dis      1.266762305  1.029387844 0.310372113         8   -1  negative   yes
## lat      1.135429569  3.084782762 0.930096998         2    1  positive    no
## crim     0.120941548  0.085181803 0.025683280         9   -1  negative   yes
## b        0.044978058  0.009116406 0.002748700        11    1  positive   yes
## tax      0.005062703  0.005459025 0.001645958         8   -1  negative   yes
## zn       0.001783381  0.005914805 0.001783381         1    1  positive    no

Variable stability can be plotted using plot_var_stability().

p1 <- plot_var_stability(fit)

# overlay directionality using colour
p2 <- plot_var_stability(fit, final = FALSE, direction = 1)

# or show directionality with the sign of the variable importance
# plot_var_stability(fit, final = FALSE, percent = F)

ggpubr::ggarrange(p1, p2, ncol=2)

By default only the predictors chosen in the final model are shown. If the argument final is set to FALSE all predictors are shown to help understand how often they are selected which is helpful when pushing sparsity in models. Original coefficients can be shown instead of being scaled as a percentage by setting percent = FALSE.

The frequency with which each variable is selected in outer folds as well as the final model is shown by as bubble size, which is helpful for sparse models or with filters to determine how often variables end up in the model in each fold.

We can overlay directionality using either colour (direction = 1) or the sign of the variable importance to splay the plot (direction = 2).

For glmnet, the direction of effect is taken directly from the sign of model coefficients. For caret models, direction of effect is not readily available, so as a substitute, the directionality of each predictor is determined by the function var_direction() using the sign of a t-test for binary classification or the sign of regression coefficient for continuous outcomes (not available for multiclass caret models). To better understand relationship and direction of effect of each predictor within the final model, we recommend using SHAP values.

Alternatively a barplot is available using barplot_var_stability().

The caret package allows variable importance to be calculated for other models types. This is model dependent. For example, with random forest variable importance is usually calculated as the mean decrease in Gini impurity each time a variable is chosen in a tree. With caret models, you may have to load the appropriate package for that model to calculate the variable importance, e.g. this is necessary for GBM models with the gbm package.

To change the colour scheme for direction = 1 overwrite scale_fill_manual().

# change bubble colour scheme
p1 + scale_fill_manual(values=c("orange", "green3"))

Explainable AI with Shapley values

The original implementation of shap by Scott Lundberg is a python package. We suggest using the R package fastshap for examining nestedcv models since it works with classification or regression as well as any model type (regression such as glmnet or tree based such as random forest, GBM or xgboost).

In the example below using the same glmnet regression model, the variable importance for predictors is measured using Shapley values. The function explain() from the fastshap package needs a wrapper function for prediction using the model. nestedcv provides pred_nestcv_glmnet which is a wrapper function for binary classification or regression with nestcv.glmnet fitted models.

library(fastshap)

# Generate SHAP values using fastshap::explain
# Only using 5 repeats here for speed, but recommend higher values of nsim
sh <- explain(fit, X=x, pred_wrapper = pred_nestcv_glmnet, nsim = 5)

# Plot overall variable importance
plot_shap_bar(sh, x)
## Variables with mean(|SHAP|)=0: tract, lat, zn, indus, age, rad

plot_shap_bar() ranks predictors in terms of variable importance calculated as mean absolute SHAP value. The plot also overlays colour for the direction of the main effect of each variable on the model, based on correlating the value of each variable against the SHAP value to see if the overall correlation for that variable is positive or negative. For regression models such as glmnet this corresponds to the sign of each coefficient. For more complex models with interactions the direction of effect may be variable and non-linear.

Note that these SHAP plots only show the final fitted model (on the whole data), whereas the variable stability plots examine models across the outer CV folds as well as the final model.

Be careful if you have massive numbers of predictors in x (see Troubleshooting).

nestedcv also provides a quick plotting function plot_shap_beeswarm for generating ggplot2 beeswarm plots similar to those made by the original python shap package.

# Plot beeswarm plot
plot_shap_beeswarm(sh, x, size = 1)
## Variables with mean(|SHAP|)=0: tract, lat, zn, indus, age, rad

size can be set to control the size of points. cex controls the amount of overlap of the beeswarm. scheme controls the colour scheme as a vector of 3 colours. Alternatively try the shapviz package.

The process for a caret model fitted using nestedcv is similar. Use the pred_train wrapper when calling explain().

# Only 3 outer folds to speed up process
fit <- nestcv.train(y, x,
                    method = "gbm",
                    n_outer_folds = 3, cv.cores = 2)

# Only using 5 repeats here for speed, but recommend higher values of nsim
sh <- explain(fit, X=x, pred_wrapper = pred_train, nsim = 5)
plot_shap_beeswarm(sh, x, size = 1)

For multinomial classification, a wrapper is needed for each class. For nestcv.glmnet models, we provide wrappers for the first 3 classes: pred_nestcv_glmnet_class1, pred_nestcv_glmnet_class2 etc. They are very simple functions and easy to extend to other classes.

For nestcv.train() models, similarly we provide pred_train_class1, pred_train_class2 etc for the first 3 classes. Again these are easily extended.

As a toy example, we show the iris dataset which has 3 classes.

library(ggplot2)
data("iris")
dat <- iris
y <- dat$Species
x <- dat[, 1:4]

# Only 3 outer folds to speed up process
fit <- nestcv.glmnet(y, x, family = "multinomial", n_outer_folds = 3, alphaSet = 0.6)

# SHAP values for each of the 3 classes
sh1 <- explain(fit, X=x, pred_wrapper = pred_nestcv_glmnet_class1, nsim = 5)
sh2 <- explain(fit, X=x, pred_wrapper = pred_nestcv_glmnet_class2, nsim = 5)
sh3 <- explain(fit, X=x, pred_wrapper = pred_nestcv_glmnet_class3, nsim = 5)

s1 <- plot_shap_bar(sh1, x, sort = FALSE) +
  ggtitle("Setosa")
s2 <- plot_shap_bar(sh2, x, sort = FALSE) +
  ggtitle("Versicolor")
s3 <- plot_shap_bar(sh3, x, sort = FALSE) +
  ggtitle("Virginica")

ggpubr::ggarrange(s1, s2, s3, ncol=3, legend = "bottom", common.legend = TRUE)

Or using beeswarm plots.

s1 <- plot_shap_beeswarm(sh1, x, sort = FALSE, cex = 0.7) +
  ggtitle("Setosa")
s2 <- plot_shap_beeswarm(sh2, x, sort = FALSE, cex = 0.7) +
  ggtitle("Versicolor")
s3 <- plot_shap_beeswarm(sh3, x, sort = FALSE, cex = 0.7) +
  ggtitle("Virginica")

ggpubr::ggarrange(s1, s2, s3, ncol=3, legend = "right", common.legend = TRUE)

One-hot encoding for categorical predictors

Binary factors generally should not have any handling problems as most functions in nestedcv convert them to 0 and 1, and directionality can be interpreted. However, multi-level factors with 3 or more levels are not clearly interpretable with SHAP values. We recommend these are one-hot encoded with dummy variables using the function one_hot(). This function accepts dataframes and can encode multiple factors and character columns, producing a matrix as a result.

Troubleshooting

explain() runs progressively slowly with larger numbers of input predictors. So if you have a very large number of predictors in x it is much better to pass only the subset of predictors in the final model. These are stored in fitted nestedcv objects in list element xsub. Or you can subset x using final_vars. For example:

sh <- explain(fit, X = fit$xsub, pred_wrapper = pred_nestcv_glmnet, nsim = 5)
plot_shap_bar(sh, fit$xsub)

sh <- explain(fit, X = x[, fit$final_vars], pred_wrapper = pred_nestcv_glmnet, nsim = 5)
plot_shap_bar(sh, x[, fit$final_vars])

Citation

If you use this package, please cite as:

Lewis MJ, Spiliopoulou A, Goldmann K, Pitzalis C, McKeigue P, Barnes MR (2023). nestedcv: an R package for fast implementation of nested cross-validation with embedded feature selection designed for transcriptomics and high dimensional data. Bioinformatics Advances. https://doi.org/10.1093/bioadv/vbad048

References

Brandon Greenwell. fastshap: Fast Approximate Shapley Values

Lundberg SM & Lee SI (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems 30; 4768–4777. http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf

Lundberg SM et al (2020). From Local Explanations to Global Understanding with Explainable AI for Trees. Nat Mach Intell 2020; 2(1): 56-67. https://arxiv.org/abs/1905.04610