Do I log my raw model inputs or human readable features?
The most common example is one-hot encoded features. Let's say one of the features of the model is STATE.
If you sent in the values as One-Hot Encoded, it would look like this
feature1 - is_CA
feature 2 - is_NY
feature 3 - is_GA
feature 50 - is_TX
While this is doable, it is cumbersome to have 50 monitors for each of the 50 one-hot encoded features. If you want to know if this feature is empty, you will get 50 monitors firing instead of just 1 monitor saying the feature STATE is empty.
Recommended: Send in the features before they are one-hot encoded.
Example: feature1 - STATE. There would be 50 possible feature values including CA, NY, GA ... etc.
How do I log vector type of features?
Native support for embeddings/vectors is coming out soon!