Tidy Modeling with R

Chapter 8. Feature Engineering with Recipes

Feature engineering entails reformatting predictor values to make them easier for a model to use effectively. This includes transformations and encodings of the data to best represent their important characteristics. Imagine that you have two predictors in a data set that can be more effectively represented in your model as a ratio; creating a new predictor from the ratio of the original two is a simple example of feature engineering.

Take the location of a house in Ames as a more involved example. There are a variety of ways that this spatial information can be exposed to a model, including neighborhood (a qualitative measure), longitude/latitude, distance to the nearest school, and so on. When choosing how to encode these data in modeling, we might choose an option we believe is most associated with the outcome. The original format of the data, for example numeric (e.g., distance) versus categorical (e.g., neighborhood), is also a driving factor in feature engineering choices.

Other examples of preprocessing to build better features for modeling include:

• Correlation between predictors can be reduced via feature extraction or the removal of some predictors.

• When some predictors have missing values, they can be imputed using a sub-model.

• Models that use variance-type measures may benefit from coercing the distribution of some skewed predictors to be symmetric by estimating a transformation.

Feature engineering and data preprocessing ...