Missing Data

In the video, I say that a machine learning algorithm won't work with missing values. This is essentially correct; however, there are a couple of situations where this isn't quite true. For example, if you had a categorical variable, you could keep the NULL value as one of the options.

Like if theme_2 could have a value of agriculture, banking, or NULL, you might encode this variable as 0, 1, 2 where the value 2 stands in for the NULL value. You could do something similar for one-hot encoding where the theme_2 variable becomes 3 true/false features: theme_2_agriculture, theme_2_banking, theme_2_NULL. You could have to make sure that this improves your model performance.

There are also implementations of some machine learning algorithms, such as gradient boosting decision trees that can handle missing values.