Building model inputs.
WHAT A FEATURE IS
A value derived from data, used as model input.
WHAT COMMON TRANSFORMATIONS ARE
Aggregations over windows Ratios and differences Encoding of categories Handling of missing values Scaling
WHAT THE HARD REQUIREMENT IS
The same computation at training and at prediction.
WHAT HAPPENS OTHERWISE
Training and serving skew: the model receives different values than it learned from.
WHY THAT IS COMMON
Training features are computed in SQL over history; serving features are computed in application code.
WHAT PREVENTS IT
One definition, used by both.
WHAT TO AVOID
Reimplementing feature logic for serving.
WHAT TO BE CAREFUL WITH
Missing value handling differing between paths Category values unseen at training Scaling parameters computed from the whole dataset
WHAT TO DO ABOUT UNSEEN CATEGORIES
Define explicit handling, rather than failing.
WHAT TO VERSION
Feature definitions, since changing one changes the model's behaviour.
WHAT TO MONITOR IN PRODUCTION
Feature distributions, against training.
WHAT DIVERGENCE INDICATES
Drift, or a broken pipeline.
