Things on this page are fragmentary and immature notes/thoughts of the author. Please read with your own judgement!

Categorical variables are very common in a machine learning project. On a high level, there are two ways to handle a categorical variable.

1. Drop a categorical variable if a categorical variable won't help the model and especially when the categorical variable has a large cardinality. User id is such an example when you build a user-level model. Of course, you can use feature hashing to reduce the dimension/cardinality of a categorical variable, and let the training process decides whether the categorical variable should be included into the model or not.

2. Encode a categorical variable. Below are some popular ways of encoding a categorical variable.

   - One-Hot Encoding
   - Label Encoding
   - Target Encoding
   - Feature Hashing
   - Weight of Evidence
   - Light G-Boost Encoding
   

   Please refer to Know about Categorical Encoding, even New Ones! and Dealing with Categorical Variables in Machine Learning for more detailed discussions. Notice that LightGBM has it's own way (Light G-Boost Encoding) of handling categorical variables. Please refer to Handle Categorical Variables in LightGBM for more discussions .

References

• Know about Categorical Encoding, even New Ones!

• Dealing with Categorical Variables in Machine Learning