- Block 6 — Transforming the data
- Objective
- build explanatory variables that carry more usable signal than
the raw columns; separate the four operations that are routinely confused —
engineering, transformation, encoding and selection; master the eight families
of construction (ratios and differences, aggregations and group statistics,
counts and frequencies, interactions, discretization, calendar and delay
features, text-derived and geospatial features); know which transformation
helps which family of algorithm and which is strictly without effect; recognize
the three failure modes specific to this work — leakage through the target,
leakage through time, and features unavailable at the moment of decision — and
place every engineered column inside the fitted pipeline so that the training
representation and the serving representation cannot diverge.
- Estimated duration
- 75 minutes
- Prerequisites
- chapters 006 (features and target), 013 (exploratory data
analysis), 014 (descriptive statistics), 017 (multicollinearity), 021
(outliers), 022 (categorical encoding) and 023 (feature scaling); the notion of
a train/test split (chapter 026) and of cross-validation (chapter 034)
- Associated quizzes
024.1-quiz-what-feature-engineering-is.md to
024.8-quiz-pipelines-and-method.md