Training-only imputation, scaling, encoding and learned state

1. Inconsistent preprocessing and the `Pipeline` fix

   scikit-learn - Common pitfalls and pipelines

   12.1 Inconsistent preprocessing with the untransformed-test MSE of 62.80 versus 0.90 after transforming, and the make_pipeline fix; 12.2.1 How to avoid data leakage (split first, fit and fit_transform on train only, transform on both)

   Done
2. `Pipeline` and `ColumnTransformer` reference

   scikit-learn - Common pitfalls and pipelines

   8.1.1 Pipeline: chaining estimators (convenience, joint parameter selection, safety: same samples train transformers and predictors) with 8.1.1.1 Usage (build, access steps, nested parameters); 8.1.4 ColumnTransformer for heterogeneous data (why preprocess inside scikit-learn: no test statistics in the preprocessors, and preprocessing parameters in the search; per-column transforms; remainder handling)

   Done
3. Numeric and categorical pipelines in the housing notebook

   Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2

   Book Chapter 2: Handling Text and Categorical Attributes and Feature Scaling and Transformation (printed pp. 73-81), then Transformation Pipelines (pp. 86-90). Notebook 02: Transformation Pipelines (num_pipeline with SimpleImputer median plus StandardScaler, ColumnTransformer with num and cat pipelines, make_column_selector variant, get_feature_names_out); Handling Text and Categorical Attributes (OrdinalEncoder versus OneHotEncoder, sparse_output flag, get_dummies mismatch, handle_unknown ignore mapping the unseen <2H OCEAN to an all-zero row); Feature Scaling (MinMaxScaler, StandardScaler)

   Done
4. Custom transformer exercise with an estimator oracle

   Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2

   Book Chapter 2: Custom Transformers (printed pp. 81-86) and Exercise 6 (p. 104). Notebook 02: Exercise 6: StandardScalerClone from scratch with inverse_transform and feature-name support, checked by check_estimator plus the allclose assertions (centered, uncentered with_mean False, round trip) and feature-name assertions

   Done

Focus on

• Fit learns state from train only: imputer statistics_, scaler mean_ and scale_, encoder categories_. Transform applies it everywhere. Section 12.1 prices forgetting the second half: test MSE 62.80 versus 0.90.
• Pipelines make the order mechanical: impute, then scale, then encode inside one object fitted once on train, with nested-parameter access for later tuning and ColumnTransformer routing per dtype.
• Unknown categories are a serving fact, not an exception. Default OneHotEncoder raises a ValueError on an unseen value such as <2H OCEAN; handle_unknown=ignore emits an all-zero row, and get_feature_names_out carries the schema into the model.
• Custom transformers must honor the estimator contract: explicit __init__ parameters, fit accepting y, learned state in trailing-underscore attributes, feature names tracked. check_estimator is the conformance oracle.

Reproduce closed-book

Traps

• Calling fit_transform on the full frame before splitting. The convenience call is the classic leak; split first, then fit_transform on train and transform on test.
• Assuming get_dummies generalizes. It silently builds disjoint columns on new categories; only a fitted encoder with a declared unknown policy survives serving.
• Forgetting the sparse default. OneHotEncoder returns sparse output by default (sparse_output True in 1.9); downstream code expecting dense breaks or densifies huge matrices, so declare the flag explicitly.

Practice

Self-test

1. Why does the wrong 12.1 run give test MSE 62.80 while the right one gives 0.90?

   Show answer Got it Missed

2. Where must SimpleImputer with median strategy be fit, and how do you check it?

   Show answer Got it Missed

3. The notebook's test frame holds <2H OCEAN, a misspelling of <1H OCEAN. What does handle_unknown=ignore emit, and what does it hide?

   Show answer Got it Missed

4. Your custom transformer takes *args for flexibility. What breaks?

   Show answer Got it Missed

Leads to

• 1.7 Baselines: constant, frequency, heuristic, last-value and simple learned Run every baseline through the same fitted pipeline.
• 1.8 Snapshots, lineage, pipeline serialization and serving parity Serialize the fitted pipeline and prove serving parity.

Comes after

• 1.3 Split design, nested selection and test isolationFit every learned preprocessing state inside the split built here.
• 1.4 Leakage: target, future, entity, duplicate, preprocessing and feedbackPut the preprocessing repair inside a pipeline.
• 1.5 Schema, provenance, missingness, duplicate and availability auditsFeed audited columns into train-only transformers with an unknown-category policy.
• 4.1 Numeric transforms and categorical encodingApply the training-only fitting rule to imputation, scaling and encoding.

Resources for this unit

• Hands-On Machine Learning with Scikit-Learn and PyTorch - Chapters 1 and 2
  Aurelien Geron (O'Reilly, 2025)primarycourse
  Read Chapters 1 and 2 of the book, then run notebooks 01 and 02 and use ml-project-checklist.md for framing, data inspection, test-set creation, preprocessing pipelines, cross-validation, error analysis and persistence. Rebuild the workflow on UCI Bike Sharing with a chronological split and prohibit casual and registered when predicting cnt.
• scikit-learn - Common pitfalls and pipelines
  scikit-learnrequiredreference
  Read Common Pitfalls sections on inconsistent preprocessing and leakage, then Pipeline and ColumnTransformer. Seed global scaling, a target-derived feature and a partition duplicate, and make tests reject each defect.