Summary

In this chapter, we covered a wide range of feature engineering techniques. We used tools to drop redundant or highly correlated features. We explored the most common kinds of encoding – one-hot encoding, ordinal encoding, and hashing encoding. Following this, we used transformations to improve the distribution of our features. Finally, we used common binning and scaling approaches to address skew, kurtosis, and outliers, and to adjust for features with widely different ranges.

Some of the techniques we discussed in this chapter are required for most machine learning models. We almost always need to encode our features for algorithms in order to understand them correctly. For example, most algorithms cannot make sense of female or male values or know not to treat ZIP codes as ordinal. Although not typically necessary, scaling is often a very good idea when we have features with vastly different ranges. When we are using algorithms that assume a Gaussian distribution...