Variable Importance via Permutation

In the previous section, we saw how to extract feature importance for RandomForest. There is actually another technique that shares the same name, but its underlying logic is different and can be applied to any algorithm, not only tree-based ones.

This technique can be referred to as variable importance via permutation. Let's say we trained a model to predict a target variable with five classes and achieved an accuracy of 0.95. One way to assess the importance of one of the features is to remove and train a model and see the new accuracy score. If the accuracy score dropped significantly, then we could infer that this variable has a significant impact on the prediction. On the other hand, if the score slightly decreased or stayed the same, we could say this variable is not very important and doesn't influence the final prediction much. So, we can use this difference between the model's performance to assess the importance of a variable...