## 5.9 Variable importance

Another important purpose of machine learning models could be to learn which variables are more important for the prediction. This information could lead to potential biological insights or could help design better data collection methods or experiments.

Variable importance metrics can be separated into two groups: those that are model dependent and those that are not. Many machine-learning methods come with built-in variable importance measures. These may be able to incorporate the correlation structure between the predictors into the importance calculation. Model-independent methods are not able to use any internal model data. We will go over some model-independent strategies below. The model-dependent importance measures will be mentioned when we introduce machine learning methods that have built-in variable importance measures.

One simple method for variable importance is to correlate or apply statistical tests to test the association of the predictor variable with the response variable. Variables can be ranked based on the strength of those associations. For classification problems, ROC curves can be computed by thresholding the predictor variable, and for each variable an AUC can be computed. The variables can be ranked based on these values. However, these methods completely ignore how variables would behave in the presence of other variables. The `caret::filterVarImp()`

function implements some of these strategies.

If a variable is important for prediction, removing that variable before model training will cause a drop in performance. With this understanding, we can remove the variables one by one and train models without them and rank them by the loss of performance. The most important variables must cause the largest loss of performance. This strategy requires training and testing models as many times as the number of predictor variables. This will consume a lot of time. A related but more practical approach has been put forward to measure variable importance in a model-independent manner but without re-training (Biecek 2018; Fisher, Rudin, and Dominici 2018). In this case, instead of removing the variables at training, variables are permuted at the test phase. The loss in prediction performance is calculated by comparing the labels/values from the original response variable to the labels/values obtained by running the permuted test data through the model. This is called “variable dropout loss”. In this case, we are not really dropping out variables, but by permuting them, we destroy their relationship to the response variable. The dropout loss is compared to the “worst case” scenario where the response variable is permuted and compared against the original response variables, which is called “baseline loss”. The algorithm ranks the variables by their variable dropout loss or by their ratio of variable dropout to baseline loss. Both quantities are proportional but the second one contains information about the baseline loss.
Below, we run the `DALEX::explain()`

function to do the permutation drop-out strategy for the variables. The function needs the machine learning model, and new data and its labels to do the permutation-based dropout strategy. In this case, we are feeding the function with the data we used for training.
For visualization we can use the `DALEX::feature_importance()`

function which plots the loss. Although, in this case we are not plotting the results. In the following sections, we will discuss method-specific variable importance measures.

```
library(DALEX)
set.seed(102)
# do permutation drop-out
explainer_knn<- DALEX::explain(knn_fit,
label="knn",
data =training[,-1],
y = as.numeric(training[,1]))
viknn=feature_importance(explainer_knn,n_sample=50,type="difference")
plot(viknn)
```

Although the variable drop-out strategy will still be slow if you have a lot of variables, the upside is that you can use any black-box model as long as you have access to the model to run new predictions. Later sections in this chapter will show methods with built-in variable importance metrics, since these are calculated during training it comes with less of an additional compute cost.

### References

Biecek. 2018. “DALEX: Explainers for Complex Predictive Models in R.” *Journal of Machine Learning Research* 19 (84): 1–5. http://jmlr.org/papers/v19/18-416.html.

Fisher, Rudin, and Dominici. 2018. “All Models Are Wrong but Many Are Useful: Variable Importance for Black-Box, Proprietary, or Misspecified Prediction Models, Using Model Class Reliance.” *arXiv Preprint arXiv:1801.01489*.